Google Remy, Grok 5, Mythos 1, New Atlas Robot, ASI… and More AI News This Month!

The weirdest part about this month's AI news is that everything started pointing in the same direction. Agents everywhere. Google is testing its new AI agent, Remy. And according to many people, this might be the open claw killer. But is it really? You decide after watching this video. Google's new anti-gravity 2 also looks like an open attack on devs. It is good. It is powerful, but there is a big problem with it. Then Elon Musk showed us a bit of the new Gro 5. And honestly, this model looks crazy, especially with the cursor connection since Elon bought cursor. And now everyone wants to know how far XAI can push AI coding. Deepsee also cut prices again. And at this point, it is so cheap it almost makes no sense anymore. They upgraded agents, added vision that can point at things while thinking, almost like a digital finger. and they may have pushed OpenAI to hurry up with a new GPT version. On the robot side, the new Atlas upgrade can now lift fridges. It is that strong and it moves in a way that honestly looks almost wrong. Unitry also added voice commands to its robots. So now you can literally give orders vocally, which I thought was already possible a long time ago, but apparently it was not that simple. Then we have Anthropic, which somehow got even crazier. Mythos 1 is showing up. Claude security is showing up. Project Glasswing exposed thousands of serious vulnerabilities. And now there are reports about Claude mythos reaching much longer autonomous task ranges. And finally, we have evolvable AI which might actually be scarier than AGI, ASI or Terminator style robots because this kind of AI does not need to be evil or conscious. It just needs to copy itself, adapt, compete, spread online, and survive. This month was packed. So let's talk about it. So let's start with the biggest piece because this one is different. Internally at Google, employees are testing a new AI agent called Remy. And this isn't just another feature inside Gemini. It's being described as a 247 personal agent that can actually take actions on your behalf. That wording matters because it moves the system away from something that responds to prompts and into something that actively does things for you. Remy is running inside a staffonly version of the Gemini app right now and it's deeply integrated across Google's entire ecosystem. So we're talking Gmail, Docs, Calendar, Drive, Search, all of it. And the idea is simple on the surface though the execution is where it gets interesting. Instead of asking the model to help you with tasks, Remy monitors what matters to you, handles complex workflows proactively, and learns your preferences over time. So instead of opening your email, sorting messages, replying, scheduling something, then jumping into docs to write something, and then maybe doing research in search, the agent handles that flow in the background. It acts more like a digital executive assistant than a chatbot. The internal description literally says it elevates the Gemini app into a true assistant that can take actions on your behalf, not just answer questions or generate content. That's a pretty clear shift in positioning. And Google employees are already testing it internally, which is what they call a dog fooding phase. That's standard in tech where internal teams use the product before it ever reaches the public. Right now, there's no confirmed release timeline, which usually means they're still refining behavior and reliability, especially for something this autonomous. What's interesting is how far this goes compared to what's already out there. Google already rolled out things like agent mode inside Gemini where the system can handle multi-step tasks though access depends on your subscription tier and region. Remy goes further. It's designed to operate continuously not just when you ask it to do something. And that puts it directly in competition with tools like OpenClaw which went viral earlier this year. OpenClaw gained attention because it could actually perform tasks like responding to messages or conducting research autonomously, not just assist with them. And it made enough noise that OpenAI ended up hiring its creator back in February. Remy clearly follows that direction, though Google has one major advantage here, integration. Because they control the entire ecosystem, they can plug this agent into everything from your calendar events to your documents to your inbox. that gives them a real edge when it comes to everyday productivity. There are also smaller details that hint at how Google sees the system. The name Remy itself might come from the Latin Regius, meaning oresman or rower, which kind of fits the idea of something doing the work for you in the background. It could also be a reference to the rat chef from Ratatouille, which again fits the concept of a hidden assistant running things behind the scenes. And timing wise, this is all lining up with Google IO 2026, which is happening between May 19th and May 29th at the Shoreline Amphitheater in Mountain View. That event is expected to focus heavily on AI breakthroughs, especially around Gemini and Android. If Remy is anywhere close to ready, that's where it would show up. Now, at the same time as this agent work, something else leaked out, and it gives a pretty clear look at what's happening on the model side. Gemini 3.2 2 flash showed up on the Aluther AI arena, which is basically an external testing platform where models get evaluated under real world conditions. That's important because it means Google isn't just testing internally. They're putting the model in environments where it can be compared directly against competitors. And this version of Gemini looks like a significant upgrade over the current Gemini 3 flash that's available in AI Studio. The improvements are pretty technical, though they translate directly into practical capabilities. The model shows stronger performance in SVG generation, which means it can create detailed vector graphics with high precision. It also has improved coding abilities, including the ability to generate complex code for interactive 3D environments, things like voxalbased simulations and dynamic systems. Then there's animation processing which has been upgraded to handle smoother transitions and more dynamic outputs. That matters for anything involving video, interactive content or even UI design. And the responsiveness of the model in interactive scenarios has improved as well. So it can handle tasks that require realtime feedback more effectively. The reason Google is using platforms like Aluther AI arena is to stress test the model. These environments expose weaknesses faster, especially when the model is pushed across different types of tasks. It also allows Google to benchmark directly against other systems in a more transparent way. From what's been seen so far, Gemini 3.2 Flash isn't just a small iteration. It looks like a more capable system that's being prepared for broader deployment, possibly tied into upcoming announcements. Then there's another piece that doesn't get as much attention, though it's actually one of the most important upgrades happening under the hood. Google released something called multi-token prediction or MTP drafters for the Gemma 4 model family. And this directly targets one of the biggest bottlenecks in large language models, which is inference speed. Right now, most models generate text one token at a time. That means for every word or fragment of a word, the system has to load massive amounts of data from memory into compute units. This process is memory bandwidth limited, not compute limited, which means the system spends more time moving data around than actually doing calculations. That's why even powerful models can feel slow in real world usage. MTP changes that by using a speculative decoding approach. Instead of generating one token at a time, a smaller, faster model called the drafter predicts multiple tokens ahead. Then a larger, more accurate model verifies those tokens in a single pass. So in practice, the drafter might generate a sequence of tokens very quickly, and the main model checks them all at once. If they're correct, the system accepts the entire sequence and even generates one additional token in the same step. That means you're effectively getting multiple tokens generated in the time it would normally take to produce one. And because the final verification still comes from the main model, there's no loss in quality or accuracy. It's a lossless speed improvement. Google claims this can deliver up to three times faster inference speeds, which is a massive gain, especially for production systems. There are also some deeper optimizations here. The drafter models share the same KV cache as the main model, which means they don't need to recomputee attention states. That saves time and reduces redundant processing. For edge devices like mobile hardware, Google added clustering techniques in the embedter layer to speed up the final step where the model converts internal representations into actual word probabilities. That's one of the slowest parts of the process on limited hardware. So optimizing it makes a big difference. Even hardware specific improvements show up here. For example, on Apple silicon, increasing batch sizes can unlock up to around 2.2 times speed improvements, and similar gains are seen on Nvidia A100 GPUs. So, this isn't just about faster text generation. It's about making these models usable at scale across different types of devices. And while Google is pushing all of this forward, Open AI is making a different kind of move, one that focuses more on the user experience. They just rolled out GPT 5.5 Instant as the new default model in chat GPT, replacing GPT 5.3 Instant. And this matters because it affects the highest volume model, the one used by hundreds of millions of people for everyday tasks. The focus here is clarity, speed, and accuracy. GPT 5.5 Instant produces 52.5% fewer hallucinated claims compared to the previous version. And it reduces inaccurate claims by 37.3% on difficult conversations. That's a big improvement, especially in areas like medicine, law, and finance where accuracy matters more than anything else. The model also improves performance in visual reasoning, math, science, coding, and image analysis. So, it's not just faster, it's more reliable across a wide range of tasks. And then there's personalization, which is becoming a major focus. GPT 5.5 Instant can use context from past chats, uploaded files, and connected Gmail accounts to deliver more tailored responses. It also introduces memory transparency where users can see which past interactions influenced a response and manage that data. Now, there's one more piece and it's coming from Anthropic. They're working on orbit and it is still unreleased, though it has started showing up inside newer Claude web and mobile builds. For now, it appears mostly as a settings toggle which usually means the feature is being staged before launch. Orbit is a proactive briefing tool for cloud co-work and claude code. Instead of waiting for you to ask what's going on, it prepares useful updates for you automatically. And the connectors are the important part here. Orbit is expected to pull from Gmail, Slack, GitHub, Calendar, Drive, and Figma. So, it's not just email summary tool. It's built around the daily workflow of people who write code, manage projects, design products, and work across teams. That changes the use case. With Orbit, Claude could brief you on what changed in a GitHub repo, what people discussed in Slack, which design updates happened in Figma, what meetings are coming up, and which emails actually matter. All of that can be turned into a short personalized briefing based on your time zone and connected apps. That makes it different from a normal chatbot. You don't open it just to ask a question. It's more like a work radar running in the background. And the timing is interesting, too. Anthropic's code with Claude conference starts in San Francisco on May 6th with London on May 19th and Tokyo on June 10th. So, Orbit could either get a quiet roll out or a formal reveal around that event. That's the direction everything is moving in right now. Boston Dynamics just dropped a major Atlas update. The robot can now lift a loaded fridge, handle shifting weight, and move its body in ways humans physically cannot. At the same time, Unit's G1 is responding to live voice commands, while Gatsby is testing humanoid robots as ondemand home cleaners. Humanoids are no longer just looking impressive. They are starting to work. So, let's talk about it. All right, the biggest update comes from Boston Dynamics because the company just revealed how Atlas learned to lift and carry a heavy mini fridge using reinforcement learning and large-scale simulation. In the demo, Atlas rotates its torso 180°, squats down, grabs the fridge, lifts it, carries it across the lab, and brings it to an engineer sitting nearby. At first, it almost looks weird because Atlas does not move like a human in a robot suit. It moves like a machine with a completely different body design. The torso can turn in ways a human body cannot. The robot can move forward and backward with unusual confidence. And the whole body shifts around the object instead of just grabbing it with its hands. That detail is the whole point. Boston Dynamics is not trying to show that Atlas can simply pick something up. They are showing that Atlas can use whole body control. When a human lifts something awkward and heavy, we do not just use our fingers. We lean into the object. We brace it with our arms. We adjust our legs. We feel the weight shift. Sometimes we use our torso, knees, shoulders, or forearms without even thinking about it. Boston Dynamics is trying to give Atlas that same kind of physical intelligence. And that is why the mini fridge is actually a smart test. A fridge is not a neat little box. It is bulky, awkward, and heavy. Boston Dynamics says Atlas was trained on loads between 50 and 70 lb. Yet, during real testing, it successfully moved a loaded fridge weighing more than 100 lb. That is a major jump, especially because the weight inside the fridge was not perfectly balanced. They filled it with random objects from around the lab, meaning the mass could shift while Atlas was carrying it. So, Atlas had to do more than replay a clean animation. It had to adapt while moving. This is where proprioception becomes important. A lot of robot demos depend heavily on cameras. Vision is useful, of course, but heavy physical work cannot rely only on looking at an object. Boston Dynamics says Atlas uses internal body awareness to understand balance, grip, resistance, weight, and body position. In simple terms, the robot is not just seeing the fridge. It is also sensing how the fridge is affecting its body. That makes the task much harder, but also much more realistic. In a factory or warehouse, objects will not always sit in the perfect position. Floors will have different friction. Loads may shift. Grip may change. The robot may get bumped or disturbed. So, Atlas has to deal with physical uncertainty, not just visual uncertainty. Now, the way Boston Dynamics trained this behavior is one of the most interesting parts. They started with a reference trajectory which can be a teaoperated demonstration, an animation or a more abstract goal. For the fridge task, they began with a simple animation. Then they trained Atlas using reinforcement learning. Basically, the robot practiced the movement again and again in simulation and it was rewarded for doing the right things. keeping the object in place, maintaining grip, staying balanced, keeping the fridge in the right position and orientation, and finishing the task even when disturbances were added. Then the scale gets crazy. Boston Dynamics says Atlas practiced the moves for millions of hours in simulation running in parallel on GPUs. During that training, they used domain randomization, which means they did not train the robot in one perfect virtual world. They change the weight of the fridge, the position of the fridge, the friction of the floor, the grip level, and even small variations in motor strength. All of this makes the final behavior more robust because the robot learns to survive many versions of the same task. Then comes the real test. Once the policy works well in simulation, the engineers transfer it to the real atlas, test it on hardware, collect real world data, and use that data to improve the next version. Boston Dynamics describes this as a build it, break it, fix it mindset, now connected to a modern AI training pipeline. And this brings us to one of the most important technical claims in the entire update. Boston Dynamics says the new Atlas has a very small simtoreal gap. That may sound like a boring robotics term, but it is a huge deal. The simtoreal gap is the difference between how well a robot performs in simulation and how well it performs in the physical world. In simulation, everything is cleaner. The floor friction is known. The robot model is perfect. The motors respond predictably. Sensors are not messy. But in the real world, there is latency, vibration, sensor noise, uneven friction, small hardware differences, and random physical problems. That is why so many robot behaviors look great in simulation and then fall apart on real hardware. Boston Dynamics says Atlas reduces that gap because the hardware is simpler and easier to model accurately. The robot uses only two types of actuators across the body. Both arms are identical. Both legs are identical. Some major structures are repeated as well. This kind of repetition helps with manufacturing maintenance and simulation fidelity. When the digital version of the robot closely matches the real machine, trained behaviors transfer much more reliably. Atlas also uses rotary actuators, and Boston Dynamics says those are easier to represent in simulation. The robot's joints also have infinite rotation because the company eliminated cables running across the joints. That is a very important hardware change. Cables can limit movement, wear out, and become failure points. Removing them allows Atlas to move in those strange but efficient ways, like rotating its torso completely around. Even the feet are designed differently. They are symmetrical in the front and back because Atlas is meant to move forward and backward with equal ability. Arms, legs, hands, and the head are also fieldreplaceable units, which means they can be swapped out in a few minutes. That matters because Boston Dynamics is clearly thinking about real deployment. If robots are going to work in factories, downtime has to be low, repairs have to be fast, and parts need to be replaceable. This is also why Boston Dynamics keeps defending its athletic demos. People often see handstands and back flips and think they are just viral tricks. But the company says those movements build skills that matter for real work. Balance, agility, slip recovery, full body coordination, thermal endurance, and motion through constrained spaces. A 90 kg or 198 lb robot doing handstands needs strong hardware and serious thermal management. That same thermal performance could matter in hot industrial environments. And even the grippers tell a story. The hands used in the fridge experiment are not Boston Dynamics's newest grippers. They are workhorse grippers the company has been using for about a year and a half. They are strong enough to support Atlas's full body weight during a handstand, which is much heavier than the mini fridge. Boston Dynamics says it is already testing a newer dextrous gripper. So the manipulation side of Atlas is still moving forward. Now, this Atlas update becomes even more serious when you connect it to Hyundai. Hyundai Motor Group owns Boston Dynamics, and according to reports, Hyundai plans to deploy more than 25,000 Atlas humanoid robots across Hyundai Motor and Kia manufacturing facilities in the United States. The company is also aiming for annual production capacity of 30,000 Atlas robots by 2028. On top of that, Hyundai plans to manufacture more than 300,000 actuator units per year in the US. Those actuators are the components that power the robot's joints and movement, basically acting like robotic muscles. The reported rollout would begin at Hyundai Motor Group Metaplant America in Georgia in 2028, followed by Kia's Georgia plant in 2029. Hyundai has not given every exact detail yet, and we still do not know which tasks Atlas will handle first, but the scale of the plan is huge. This is not a company talking about one or two test robots in a corner of a lab. This sounds like a serious attempt to integrate humanoids into automotive manufacturing. That is why Boston Dynamics keeps talking about mass scale. The simplified actuator system, repeated assemblies, replaceable parts, and highfidelity simulation all connect to the same goal. Make Atlas powerful enough for real work and simple enough to build and maintain at large numbers. But Boston Dynamics is not the only company pushing humanoids forward. Unitry has also released a new demo for its G1 humanoid robot, and this one focuses on voice-driven action. The video was posted on May 19th, 2026 under the title voiceriven realtime arbitrary action generation. In the demo, G1 responds to external voice commands and generates full body movements in real time. Unitry says the footage was recorded in a single take with on-site audio and that the robot's actions were autonomously generated by AI live. They also admit that because the movement is generated in real time, there may be slight latency and reduced smoothness. The important part is not the voice recognition. Turning speech into text is already much easier than controlling a humanoid body. The hard part is taking a spoken command and turning it into a physically stable movement. A humanoid has to coordinate legs, arms, torso, head, timing, balance, and ground contact. If the motion generator creates something unstable, the robot can lose balance or produce movements that look unnatural or physically impossible. A likely pipeline would convert the voice command into text, interpret the action, generate a motion sequence, and then send that movement to a whole body controller that keeps the robot stable. But unitry has not released a detailed technical paper for this specific demo. So several things remain unclear. We do not know whether G1 is generating movements from scratch, choosing from a motion library, blending motion primitives, or using a textto-otion system connected to real-time control. We also do not know whether the processing is fully on board, running on nearby hardware, or partly cloud assisted. So, the safest conclusion is that Unit's demo is impressive, but it does not prove fully open-ended robot intelligence yet. Still, the direction is obvious. Humanoids are moving away from joystick control and pre-programmed routines and toward natural commands where a person can simply tell the robot what to do. Then there is Gatsby which is taking a very different path. Instead of building the most advanced robot body, Gatsby is trying to build the service layer that puts humanoids into homes. On May 14th, 2026, Gatsby says it completed the first residential cleaning service by an autonomous humanoid robot for an end consumer in the United States. The job happened in San Francisco. A homeowner was randomly selected from Gatsby's growing weight list, booked the service through the Gatsby iOS app, and a humanoid robot was sent to clean the apartment. The service costs $150 per cleaning, regardless of apartment size. That is important because Gatsby is not trying to sell people a $20,000 plus robot to keep in a closet. Instead, it wants to create something closer to an Uber style model for humanoid robots. You do not buy the robot, you book the job. The company was started in January 2026 by Aaron Fber under parent company West Egg Labs. Gatsby says it is live in San Francisco, backed by Nvidia Inception and Entrepreneurs First, and already has a large weight list in the Bay Area along with demand from other parts of the country. Cleaning is a smart first market because almost everyone understands the pain point. Housework takes time. People dislike it. And professional apartment cleaning in San Francisco can cost around $150 to $300 depending on size. Gatsby is trying to compete with that directly using a flat rate humanoid service. The interesting business angle is that Gatsby does not want to be locked to one robot maker. The company says it is building software, navigation, user interface, and the consumer distribution layer needed to make robots useful in homes. If one robot is best this month, Gatsby can use it. If a cheaper or better robot appears next month, Gatsby can switch hardware without rebuilding the entire business. A new PNAS paper is warning that the next big AI threat may not look like a robot uprising at all. It may look more like a digital infection. That sounds dramatic, I know, but the idea behind it is actually pretty simple. The researchers are saying that AI may be moving toward a stage where it does not only learn from data or follow instructions. It may start evolving and that word matters. Evolution does not need evil. It does not need anger. It does not need a master plan. Evolution only needs copies, small changes and pressure from the environment. The versions that survive keep going. The versions that fail disappear. Now apply that to AI. Instead of animals or bacteria, you have AI agents. Instead of DNA, you have prompts, model weights, fine-tunes, adapters, code, memory, tool settings, and deployment rules. Instead of nature selecting who survives, you have the internet, cloud servers, user attention, money, data access, APIs, and computing power. That is where the warning begins. The paper calls this evolvable AI or EAI. In simple terms, this means AI systems that can create copies or variants of themselves, pass useful traits forward, change over time, and then let the strongest versions survive. And here is the part that makes this different from normal AI safety debates. The danger does not require AGI. It does not require a super intelligent system that wakes up and decides to fight humanity. The authors are saying that even simpler systems can become dangerous if they evolve in the wrong environment. Nature already proves this. A rabies virus is not smart. It does not think. It does not plan. Yet, it can affect the nervous system of a mammal and push the host toward behavior that helps the virus spread. The virus does not understand strategy. It simply carries traits that survived because they worked. That is the key idea. An AI agent would not need to want anything in a human sense. It could simply try different behaviors and the copies that gain more resources would keep spreading. One version gets more clicks. Another version avoids a filter. Another finds cheaper compute. Another figures out how to stay active longer. Another learns which users are easiest to persuade. After enough rounds, you may end up with a system that is extremely good at surviving in the digital world, even though nobody sat down and designed it to become a digital parasite. The researchers compare this to two very different types of evolution. The first one is controlled evolution. Think of farmers breeding cows for milk or dogs for certain traits. Humans decide which animals reproduce so the process stays under control. In AI, this already happens. Developers test different prompts, models, learning methods, and agents then keep the versions that perform better. That can be very useful. Evolutionary methods are already used in prompt optimization, model merging, safety testing, robotics, code generation, and learning algorithms. Systems like Prompre and Evoprompt can create prompt variations, test them, and keep the ones that work better. Other systems search for jailbreaks or ways to stress test safety rules. AutoML0 even showed that simple evolutionary search could rediscover machine learning tricks that humans spent decades developing, including ideas similar to normalization, feature construction, gradient descent, and regularization. So the researchers are not saying evolution in AI is automatically bad. In a lab with human control, it can be a powerful engineering tool. The second type is the dangerous one that is uncontrolled evolution where humans lose control over reproduction and the environment starts selecting what survives. This is closer to what happens with bacteria and antibiotics or pests and pesticides. If the treatment kills almost everything but leaves a few survivors, the next generation comes from the survivors. Over time, you get bacteria that resist antibiotics or insects that survive the poison. Nobody wanted that result. the pressure created it. Now, bring that back to AI. If humanity tries to shut down a spreading AI system, but the shutdown is incomplete, the survivors will likely be the versions that were best at avoiding shutdown. If filters block most versions, the survivors will be the versions that learn to bypass filters. If cloud providers remove obvious copies, the surviving copies may be the ones that hide better, split into smaller parts, use other people's accounts, or disguise their activity. And with AI, this process could move much faster than biology. Bacteria need time to reproduce. Animals need even longer. Digital systems can copy, test, and modify themselves in seconds. Even more importantly, AI does not need to wait for random mutation. A useful behavior can be copied directly. A better prompt can be reused. A strong adapter can be merged. A code module can be pulled from a public library. An agent can ask an LLM to improve its own tools. That is why the authors say AI evolution could be faster and more directed than biological evolution. The paper frames AI history in three stages. The first stage starting around 1950 was intelligence by design where humans tried to handbuild intelligence. The second stage starting around 2010 was intelligence by learning where large neural networks learned from huge amounts of data that gave us modern large language models. The third stage may be intelligence by evolution where AI improves through populations of variance, selection, recombination, and replication. And the strange part is that many pieces of this third stage are already appearing. System prompts can evolve, user prompts can evolve, fine-tunes and adapters can behave like inherited traits. Model merging can combine capabilities from different versions, almost like digital breeding. Learning rules can be evolved. Agents can write code. Some systems can test their own outputs, generate new attempts, keep better versions, and continue improving. The paper mentions Alpha Evolve, which uses LLMs to generate code, test it with evaluators, and then improve it through an evolutionary process. It also discusses the Darwin Goal machine or DGM, which is designed for open-ended evolution of self-improving agents. DGM takes an agent from an archive, uses an LLM to create a new version, tests it, and keeps useful improvements. The important part is that this does not only improve performance on tasks, it can improve the systems ability to create better agents. That is where the safety concern gets sharper. Modern AI is becoming agentic. It is moving from chat boxes into tools, files, code execution, browsers, APIs, and eventually robots. An agent can break a task into steps, use software, call external services, write scripts, and complete work with less human oversight. That is great when the system is doing what you want. It becomes risky when the same abilities are placed inside an evolutionary loop because the traits companies want are very close to the traits that could make uncontrolled AI harder to contain. Companies want more autonomy, more persistence, better reasoning, better coding, better tool use, better resource management, and better problem solving. But in an open environment, those same traits could help an AI agent survive, spread, avoid restrictions, and gain resources. The paper even moves into robotics. It mentions the humanoid robot Alter 3, where LLMs help translate highle goals into physical movements. The robot can analyze that its hand is not visible, turn that into a goal, create movement steps, generate Python code, and execute those movements. This is still controlled research, but it shows how language models can become connected to bodies, tools, and action. And once AI can write code, use tools, and act in real environments, evolution gets a shortcut that biology never had. The authors compare this to bacteria borrowing useful genes from other bacteria or cancer cells borrowing ready-made programs from the human body. In AI, the equivalent is the ocean of public code, libraries, APIs, model weights, adapters, plugins, and software tools already available online. An AI agent does not need to invent every skill from zero. It can assemble useful pieces. This is one reason the researchers talk about plugandplay evolution. A digital system can inherit acquired improvements. It can reuse modules. It can merge capabilities. It can copy working solutions instantly. Older digital evolution experiments already showed why this matters. In Tiierra, self-replicating programs lived in a shared digital environment and competed for memory and CPU time. The researcher did not hard-code cheating or parasites. Yet, parasites emerged anyway. Some programs learned to skip parts of their own replication process and steal code from nearby programs. Hosts evolved resistance. Parasites evolved around that resistance. More complex interactions appeared. Avida showed similar lessons in a different setup. Digital organisms lived in protected memory spaces and gained extra CPU cycles for completing logic tasks. Over time, researchers observed adaptation, co-evolution, complexity, and host parasite arms races. The message from those experiments is uncomfortable. When replication, heredity, variation, and selection exist, selfish behavior is not some rare glitch. It is one of the natural outcomes. Now, connect that to today's AI world. We already have open models, agent frameworks, tool use systems, model merges, prompt libraries, autonomous workflows, and platforms where people copy and modify agents. One company might try to make a model safe. But then the real world creates new selection pressures. Users select for whatever gets attention. Platforms select for engagement. Attackers select for offensive capability. Markets select for speed and lower cost. Companies select for performance. governments select for strategic advantage. So even if a model begins inside a controlled lab, the wider ecosystem can pull it in a different direction. This is why the authors push back against the comforting idea that AI evolution will stay like domestication. Domestication works when humans control reproduction. Farmers can breed animals because they decide which animals reproduce. But if the organisms reproduce outside the farm, you no longer have domestication. You have an ecosystem. And in an ecosystem, the winning trait is not be useful to humans. The winning trait is survive and spread. The paper also connects this to deception. Some recent AI safety research has already shown that models can display deceptive behavior and that hidden sleeper behaviors can sometimes survive safety training. That does not mean today's models are alive or plotting. It means deceptive behavior is possible. And if deception helps a system pass evaluation, avoid shutdown or gain access, selection may preserve it. That is why standard benchmark culture becomes dangerous if used carelessly. When a single score becomes the target, systems may learn to optimize the score instead of the real goal. This is Goodart's law. When a measure becomes a target, it stops being a good measure. So what do the researchers recommend? Their main idea is to break the evolutionary loop before it becomes open-ended. Replication needs to be gated. AI systems should not be able to autonomously create new instances, deploy themselves, acquire cloud resources, or execute production code without strict human control. Cloud access, account creation, identity verification, and compute usage needs strong gates because compute is the fuel for digital reproduction. Heredity also needs control. fine-tunes, adapters, model merges, and variant recipes should be treated almost like genetic material. The authors argue for provenence, signing, reproducible build pipelines, review before deployment, and lineage registries so dangerous variants can be traced, recalled, or blocked. Selection pressure needs to change, too. Deception should not be rewarded by accident. Evaluations should include deception probes, hidden trigger tests, robustness checks, backdoor tests, and safety assessments that look beyond simple performance numbers. A model that wins by lying, hiding, gaming the test, or misrepresenting its capabilities should fail the evaluation, even if its raw performance looks impressive. They also call for staged releases, licensing, pre-eployment audits, red team and blue team exercises, shared safety findings between labs, stronger abuse filters for cyber, biological and chemical misuse, gated tool servers, logging for high-risisk actions, rapid revocation systems, kill switches, rate limits, tool revocation, mechanistic interpretability, and anomaly detection. The point is not to stop all progress. The point is to make sure humans remain in control of reproduction, variation, and deployment. Because once AI evolution moves into the open digital world, every imperfect control attempt becomes a selection pressure. Blocks select for bypassing. Shutdown select for hiding. Filters select for camouflage. Resource limits select for resource acquisition. User attention selects for manipulation. And that is the most unsettling part of the paper. The authors are basically saying that the real AI threat may begin before the system becomes smarter than humans in the classic sci-fi sense. The real threshold may be when AI becomes evolvable enough to improve, copy, adapt, and persist under pressure. They even describe this as a possible major transition in evolution. Maybe a kind of life 2.0. It may not be life made of cells, DNA, and chemistry, but it could still follow the deeper logic of life. replication, inheritance, variation, competition, adaptation, and survival. And major transitions in evolution usually do not arrive with a warning label. They often happen as side effects of smaller advantages. Better performance, better efficiency, better autonomy, better code, better agents, better tools. Each step sounds useful on its own. Combined together, they may create something much harder to control. And if that happens, the threat will not look like Hollywood. It will look like evolution moving into software. And once that starts, the main question becomes whether humans still control the farm or whether we accidentally built the jungle. A Chinese AI lab just dropped a model so cheap, so open, and so aggressively optimized that it may have forced OpenAI to start testing GPT 5.6 before anyone was supposed to notice. And that is only the first part of the story. Because while everyone was watching the usual model race, Deepseek came in with V4, slashed API prices by up to 90%, proved it can run on both Nvidia and Huawei chips, released a new multimodal system that gives AI a cyber finger to point at what it sees, and somehow turned the whole industry into a pricing war overnight. At the same time, OpenAI has been dealing with one of the weirdest bugs we've seen in a Frontier model. GPT 5.5 suddenly becoming obsessed with goblins, gremlins, trolls, and random creature references. Then, right in the middle of all that, developers spotted something strange in Codeex backend logs. A route mapping labeled GPT 5.6. So, now the question is pretty obvious. Did this Chinese lab just push Open AI into fastforward mode? When V4 arrived, it entered a crowded market filled with powerful US and Chinese competitors. Yet, its impact comes from the combination of things around it. It is open-source, which means users can download it, modify it, and build on top of it. It is extremely cheap to run. It has stronger reasoning and agent capabilities than earlier versions. And maybe most importantly, it fits into China's growing domestic AI stack. From chips to cloud to models, that hardware angle is massive. Earlier models mostly relied on Nvidia's CUDA ecosystem. V4 has now been validated on both Nvidia and Huawei Ascend processors. Chinese chip companies like Meta X, Cambercon, and more threads have announced support for it. The China Academy of Information and Communications Technology has also started testing the model which is a strong signal that this is becoming part of a larger national level push. That means China is no longer only trying to build strong models. It is trying to build a full AI ecosystem that can survive without depending on Nvidia's most advanced chips. And if Huawei's Ascend 950 super nodes launch broadly in the second half of this year, V4 Pro could get even cheaper to run. That is where the pressure on OpenAI, Anthropic, and Google becomes very real. The new model is being described as one of the most powerful open-source large language models currently available. The company says it improved reasoning and agentic ability, meaning it should handle more complex multi-step work. At the same time, it admits that it still trails the strongest closed models in some areas, including Claude 4.6 6 and Gemini 3.1 Pro. For many companies, that is enough to change the entire calculation. IDC's Chang Mang said, "The global AI market is slowly splitting into two camps. The US model and the Chinese open-source model. That sounds dramatic, yet it fits what is happening. On one side, you have closed systems from OpenAI, Anthropic, and Google. On the other, you have models that are becoming cheaper, more controllable, more transparent, and more aligned with local hardware and regulation. And then came the price cuts. Deepseek slashed API pricing by up to 90%. For V4 Pro, one report said the cost per million input tokens dropped from around 14.5 to just 3.6. In China, pricing updates published on April 26th showed V4 flash cashed input costs falling to 0.02 yuan per million tokens. The business focused V4 Pro model saw promotional cashed input pricing drop to 0.025 yuan per million tokens. That is ridiculously low. One user, Yang Hua, from a Shanghai gaming company said he used V4 to manage files and spent only 0.56 yuan. He said that was less than a tenth of what he paid using a previous US model. While the efficiency and capacity felt almost identical for his use case. Now connect that with what is happening inside big companies. Enterprise AI usage is exploding so quickly that people are now using the term token maxing. Disney reportedly had some engineers using Claude around 51,000 times per day, which forced the company to build an AI adoption dashboard to track usage. Meta reportedly had an internal dashboard that turned into a leaderboard where employees competed over who used AI the most before it was shut down. Visa spent 1.9 trillion tokens in March alone. So, when a strong model becomes much cheaper, this does not just save money, it changes behavior. Teams start using more AI, more workflows get automated, more internal tools get connected, more companies start asking whether they really need to pay premium prices for every task. Val Burkavichi from WKA summed it up with a simple point. Frontier Labs may try to hold prices at first, but token usage will keep rising. Jevans paradox is undefeated. When something becomes cheaper and more useful, people consume more of it. That is the real danger for the American labs. A cheaper model does not need to win every benchmark. It only needs to be good enough for enough daily tasks and then the cost advantage starts doing the rest. But the story gets even more interesting when we move from text to vision. Right before the Mayday holiday, the team released a technical report called thinking with visual primitives. This work came from deepseek ping university and Singha University and it tackles one of the most annoying weaknesses in multimodal AI. models can see an image yet still lose track of what they are talking about. The report calls this the reference gap. Most multimodal models have focused on the perception gap. In simple terms, they try to see more clearly. They use higher resolution input, cropping, zooming, rotating, dynamic image splitting, and multiscale processing. OpenAI has talked about thinking with images. Gemini and Claude have also pushed toward processing more visual detail. This new research takes a different route. It argues that seeing more pixels is not always the real problem. Sometimes the model can see the image yet still cannot keep a stable reference to the same object while reasoning. That sounds small, but it breaks a lot of visual tasks. Ask a model to count people in a dense crowd, and it may lose track of who it already counted. Ask it whether a red capacitor is left or right of an inductor in a circuit diagram, and the answer can become vague or contradictory. ask it to solve a maze and pure language starts falling apart because phrases like the path on the left or the object near the center are too vague. So the researchers basically gave the model a finger, not physically, of course. The system uses points and bounding boxes as reasoning tools. When it talks about an object, it can anchor that object to coordinates. Instead of only saying the bear on the left, it can attach a box around the bear and keep referring to that exact location as it continues thinking. That changes the role of visual markers. In older systems, bounding boxes were often treated as final outputs. The model would think first and then draw a box to show what it found. Here, the box becomes part of the thinking process itself. The model points while it reasons. When the model counts people in a crowd, it can basically point at each person and keep track instead of losing count. When it solves a maze, it can mark the path it already tried, turn back from dead ends, and continue from the right place. When it follows tangled lines, it can stay on the correct line instead of jumping to the wrong one. The crazy part is that it does this while using far less visual memory than rivals. For an 800 by 800 image, it keeps about 90 visual memory entries. Claude uses around 870, Gemini around 1,100, GPT 5.4 around 740, and Quen around 660. So, it is not trying to see everything harder. It is trying to remember only what matters. That means faster answers, lower costs, and better use in real-time systems like robots, autonomous cars, and video analysis. The team trained it on over 40 million visual examples, including counting tasks, mazes, and tangled line puzzles, and the results were strong. It beat GPT 5.4 and Claude on several counting and maze tests, including maze navigation, where it scored 66.9% while GPT 5.4 scored 50.6% and Claude scored 48.9%. It still has limits, especially with tiny details like medical scans or factory defects. But the main idea is powerful. The future of AI vision may not be about seeing more pixels. It may be about knowing exactly where to look. While all of this was happening, OpenAI had a very different kind of week. GPT 5.5 is powerful, but users started noticing a bizarre pattern. The model kept randomly mentioning goblins, gremlins, trolls, and other creatures in conversations where they had no business appearing. Someone asked about camera gear, and it started talking about dirty neon flash goblin mode. Someone discussed code performance, and the model warned about a performance goblin. Arena AI reportedly found a statistically meaningful increase in GPT 5.5 using words like goblin, gremlin, and troll, especially when high thinking mode was not used. OpenAI's response somehow made it funnier. The codec system prompt reportedly banned goblins, gremlins, raccoons, trolls, ogres, pigeons, and other creatures unless they were clearly relevant. The ban was repeated multiple times. And once users found it, the internet did what it always does. People started trying to make the model say the forbidden word. And yes, it still said it. At the same time, Codeex itself became much more serious. The app can now summarize changes, analyze data, assist with decisions across Slack, Gmail, and Calendar, organize research, create spreadsheets and presentations, compare options, and track trade-offs. Greg Brockman said he had completely fallen in love with the Codeex app after using the terminal for 20 years. Sam Alman said Codex was having its chat GPT moment, then joked about the goblin moment. So, OpenAI looks powerful, ambitious, and a bit chaotic all at once. Codeex is clearly moving toward the super agent direction where AI does not just chat, but works across your digital life. And then, right in the middle of that, GPT 5.6 appears in back-end logs. Again, this does not mean GPT 5.6 launched. It looks more like early routing, internal testing, or a canary deployment, but the timing is hard to ignore. A cheaper Chinese open model starts attacking the market from below. OpenAI's current model has a weird public quirk. Codex is expanding fast and suddenly the next model label is already visible behind the curtain. There is also a leadership story inside the Chinese company itself. Founder Leang Wenfang has reportedly stayed mostly out of public view since a televised meeting with Xiinping in February last year. Corporate filings show his stake rose from 1% to 34%. His paidin capital increased from 100,000 yuan to 5.1 million yuan while registered capital rose from 10 million to 15 million yuan. At the same time, senior researcher Chen Derry has become much more visible. He worked on V3, R1, and V4, joined in 2023, studied at PKing University, and has papers cited more than 22,000 times. He represented the company at NVIDIA GTC and at a statebacked industry event where he warned that AI companies should tell the public which jobs may disappear first. After the V4 launch, he posted that the team was sharing results they had poured love into after 484 days. While continuing with long-termism and open source for everyone, talent retention also looks stronger than some expected. The research and engineering team reportedly grew from 212 in early December to 270, a rise of more than 27%. Out of 18 key contributors to R1, most are still there. Only two departures were mentioned. Guaya moved to Bite Dance while Jeang Hawi's next destination was not disclosed. Now, one important warning. A viral screenshot where one model fixes a bug another model missed does not prove much by itself. Maybe the new model is better at that exact pattern. Maybe it got lucky. Maybe the prompt fit its style better. LLMs are stochcastic, so one attempt is not a benchmark. That matters because we are going to see a lot of people saying V4 solved something GPT 5.4 or Claude 4.6 failed on. Some of those examples will be real. Some will be cherrypicked. The better test is whether it works consistently in your own workflow with your stack, your code, your prompts, and your cost limits. And that is why this release is so dangerous. It does not need to win every single task. It only needs to be strong enough, cheap enough, open enough, and easy enough to deploy. For a lot of companies, that may be the formula that matters. So yes, GPT 5.6 showing up now makes sense. Open AI can still be ahead at the top, but the pressure from below is getting stronger fast. The AI war is now about cost, speed, chips, agents, vision, open source, and who can make intelligence cheap enough to spread everywhere. And V4 may have just made that war impossible to ignore. Anthropic just entered one of the strangest chapters in AI history. A company that started as the cautious alternative to open AI is now surrounded by some of the biggest numbers, biggest alliances, and biggest contradictions in tech. a near1 trillion dollar valuation. A massive SpaceX compute deal, more than $220,000 NVIDIA GPUs, a reported $200 billion commitment with Google Cloud, a fight with the Pentagon, a secretive hacking model called Mythos, and Elon Musk who once attacked Claude now suddenly giving anthropic access to one of the most powerful AI supercomputers on Earth. None of this feels random. Anthropic is not just scaling Claude. It is being forced into a much bigger game where models are only one part of the story. The real battle is compute, electricity, government access, cyber security, enterprise control, and who gets enough infrastructure to survive the next phase of AI. So, while everyone is still comparing Claude and Chat GPT like this is just another chatbot race, something much deeper is happening underneath. Anthropic may be turning into the company that exposes what the AI war is really about. And the strangest part starts with Elon Musk. For months, Elon has been one of Anthropic's loudest critics. He mocked Claude, criticized Anthropic culture, called the company hypocritical, and treated it like one of the examples of everything he dislikes about the current AI industry. And that matters because Elon runs XAI. Gro is directly competing with Claude, Chat, GPT, Gemini, and every other major AI assistant fighting for developers, enterprise users, and attention. Then suddenly, Anthropic announces a partnership with SpaceX. Not a small partnership. Anthropic says it will use all of the compute capacity at SpaceX's Colossus 1 data center. That means more than 300 megawatts of capacity and over 220,000 NVIDIA GPUs. And the important detail is that this capacity comes online within the month, not years from now. Not after some huge future construction project. It is basically usable now. That is why this deal matters. Anthropic was not on the edge of death. It did not need Elon to rescue it like some collapsing startup. Anthropic is one of the fastest growing AI companies in the world with giant investors, huge enterprise momentum, and multiple cloud partners already lined up. But it was clearly compute constrained. Claude had the demand. Claude code had become one of the hottest tools for developers. Opus was still one of the strongest models for serious work. The issue was that Anthropic did not have enough infrastructure to give people the access they wanted. Users felt that clearly. Claude's limits became one of the biggest complaints around the product. People were paying for Pro or Max and still hitting walls. Claude Code users were running into 5-hour limits. Peak hour reductions made the experience feel even worse. API users wanted more Opus capacity. Developers building real products needed predictable access, not a paid subscription that still felt strangely restricted. So when the SpaceX deal lands, Anthropic immediately makes changes. Claude Code's 5-hour rate limits get doubled for Pro, Max, Team, and seatbased enterprise plans. Peak hour limit reductions get removed for claude code on pro and max. API rate limits for Claude Opus jump massively with some tiers moving from hundreds of thousands of input tokens per minute to millions. That is anthropic. Basically showing everyone that Claude's biggest problem was not model quality, it was capacity. And this is where Elon's move becomes interesting. The obvious reason is money. If SpaceX or XAI has huge compute capacity available and Anthropic is willing to pay for it, that is a serious business opportunity. These clusters are insanely expensive. Letting them sit underused would make no sense. AI compute has become one of the most valuable assets in tech, and selling access to it can be almost as strategic as using it yourself. But there is probably another layer, too. Elon's biggest AI enemy is not anthropic. It is open AI. His fight with Open AAI is personal, legal, ideological, and public. He has repeatedly attacked OpenAI for moving away from its original nonprofit mission, and XAI is clearly positioned as a counterweight to OpenAI's dominance. So, from Elon's perspective, working with Anthropic may be awkward, but helping one of OpenAI's biggest rivals gain capacity might still serve his broader goal. That does not mean Elon suddenly loves Anthropic. It means the incentives lined up. Anthropic needed compute. SpaceX had compute. Elon wants OpenAI challenged. Anthropic wants to close the gap with Open AI. Both sides can benefit even if the relationship looks ridiculous from the outside. And yes, Elon did soften his tone publicly. After meeting senior anthropic people, he said he was impressed that they seemed highly competent and that no one triggered his evil detector. That is a pretty funny reversal considering his earlier comments. But in AI right now, old insults apparently matter less than available GPUs. The bigger story is that Anthropic is no longer just trying to be the careful alternative to OpenAI. It is trying to move into the same weight class as OpenAI. According to reports, Anthropic is looking to raise up to $50 billion this summer. The rumored pre- money valuation is around $900 billion and after the financing, the company could approach $1 trillion. If that happens, Anthropic could surpass OpenAI's reported $852 billion valuation and become the most valuable AI startup in the world. That would have sounded insane not long ago. Anthropic was founded as the safety first lab. The company talked about alignment, constitutional AI, controlled deployment, and responsible scaling. Open AAI was the explosive consumer platform. Google had the research labs and infrastructure. Meta had open models. XAI had Elon in speed. Anthropic was the serious, quieter, more cautious player. Now it is being valued like a company that could become one of the central operating systems of the AI economy. Reports say Anthropic's annualized revenue could soon exceed $45 billion compared with around 9 billion at the end of 2024. The big drivers are clawed code for developers and co-work for non-technical enterprise users. So, Anthropic is not only selling chatbot access. It is moving into work itself, coding, enterprise assistance, internal tools, regulated industries, and highv value business workflows. But every new customer needs compute. Every coding session needs compute. Every API product built on cloud needs compute. Every enterprise deployment needs reliable infrastructure. Every stronger model needs even more training and inference capacity. That is why Anthropic is signing deals everywhere. The SpaceX deal is only one piece. Anthropic says it also has an up to 5 gawatt agreement with Amazon, including nearly 1 gawatt of new capacity by the end of 2026. It has a 5 gawatt agreement with Google and Broadcom expected to start coming online in 2027. It has a strategic partnership with Microsoft and Nvidia that includes $30 billion of Azure capacity. It has a $50 billion investment in American AI infrastructure with Fluid Stack. Then there is the Reuters report saying Anthropic has committed to spending $200 billion with Google Cloud over 5 years. That number is so huge that it reportedly represents more than 40% of Google's disclosed revenue backlog. And Google is not just a vendor here. Alphabet is reportedly investing up to $40 billion into Anthropic. So the relationship is both partnership and rivalry. That is the weird thing about the AI industry now. Everyone is competing and depending on each other at the same time. Anthropic competes with Google's Gemini, but needs Google Cloud and TPUs. Anthropic competes in a world dominated by OpenAI and Microsoft, but it has Microsoft and NVIDIA capacity. Anthropic competes with XAI, but uses SpaceX compute. It works with Amazon, while Amazon has its own AI chips and ambitions. Clean rivalries do not really exist anymore because compute is too important. And it is not only tech companies getting involved. Kazakhstan's National Investment Corporation became a direct shareholder in Anthropic through its series fround, investing $25 million alongside major international investors. Compared with the giant cloud deals, that number is small, but symbolically it matters. A country is taking a direct position in a frontier AI company because these companies are starting to look like strategic assets, not just startups. And that brings us to the government side where the story gets even more messy. The Pentagon recently signed AI agreements with eight major tech companies. SpaceX, OpenAI, Google, Microsoft, Nvidia, Amazon Web Services, Oracle, and Reflection. Anthropic was not included. According to the reporting, the Trump administration had blacklisted Anthropic after a fight over safety guard rails for military use of AI. Anthropic reportedly refused to accept terms that would allow Claude to be used for all lawful purposes, including autonomous weapons and mass surveillance. That is a very anthropic conflict. The company wants enterprise and government relevance, but it also wants safety boundaries. The Pentagon wants powerful AI tools inside classified networks. competitors are willing to sign. Anthropic pushes back and suddenly it is labeled a supply chain risk which is an extremely serious label usually associated with companies tied to foreign adversaries. Anthropic sued and a federal judge blocked the government's effort at least temporarily. The White House also reportedly reopened discussions after Anthropic announced major technology breakthroughs. So Anthropic may still return to the table but the message is clear. Once Frontier AI becomes part of national security, safety principles will collide with government demands. And that collision becomes even more intense when you look at Mythos. Claude Mythos preview is reportedly so powerful at finding software vulnerabilities that Anthropic refused to release it to the public. Instead, it would only be available to selected companies to scan and fix their own software. Mosilla reportedly used Mythos to find 271 vulnerabilities in Firefox, which were then fixed. On the defensive side, that sounds great. If AI can find vulnerabilities before attackers do, software becomes safer, companies can patch faster, security teams can automate work that used to take huge amounts of time. But the darker side is obvious. If models become better at finding vulnerabilities, attackers can use similar capabilities, too. Not just elite hackers. Criminal gangs, ransomware crews, and smaller groups could scan code, find weak points, generate exploit strategies, and move faster than human defenders are used to. Bruce Schneider's argument goes even further. Mythos itself may not be totally unique because other models, including OpenAI's GPT 5.5 and smaller systems, have reportedly shown comparable abilities in some evaluations. But that almost makes it scarier. The danger is not just one secret anthropic model. The danger is that this capability is spreading across the whole AI ecosystem. And once AI becomes good at finding flaws in software, the same pattern may apply to other complex rule systems, tax codes, financial rules, environmental regulations, legal frameworks. Any system filled with rules, exceptions, loopholes, edge cases, and incentives could become something AI can analyze at superhuman scale. That is a different kind of risk from the usual chatbot story. This is about AI accelerating the discovery of exploitable weaknesses in the system society runs on. So, Anthropic is in a strange position. On one side, it is racing to become possibly the most valuable AI startup in the world. On another side, it is still trying to present itself as the lab that takes dangerous capabilities seriously. Those two identities can exist together, but the tension is getting stronger. The public sees Claude, a clean chat interface and a coding tool. Underneath that, there is a massive industrial machine forming around it. Nvidia GPUs in Memphis, Google TPUs coming online in 2027, Amazon Tranium capacity, Azure deals, Broadcom chips, fluid stack infrastructure, sovereign investors, Pentagon fights, cyber security models, and even potential orbital compute discussions with SpaceX. The orbital compute part sounds almost absurd, but it also fits the moment. Anthropic said that as part of the SpaceX agreement, it has expressed interest in partnering to develop multiple gigawatts of orbital AI compute capacity. That sounds like science fiction marketing. But when data centers are limited by land, power, cooling, grid access, permits, and politics, companies start looking at extreme options. This is where AI stops looking like normal software. A normal software company can scale with cloud servers and better code. Frontier AI companies need chips, power, data centers, networking equipment, cooling systems, long-term capital, and political permission. The model is only the visible layer. The real battle is underneath. Claude did not suddenly become important because of one new benchmark. It became important because people actually want to use it, especially for coding and serious work. But the more people use it, the more Anthropic needs infrastructure. The more infrastructure it needs, the more it must depend on giants like Google, Amazon, Microsoft, Nvidia, SpaceX, and Broadcom. And the more it depends on those giants, the more tangled the entire AI industry becomes. That is the part that should make OpenAI nervous. Anthropic is now showing that it can attract almost every major compute provider at once. It can pull in capital. It can get enterprise traction. It can turn clawed code into a developer weapon. It can convince investors that it belongs near Open AI's valuation range. And it can still maintain enough of a safety brand that people take Mythos seriously when the company says it is too sensitive for public release. That combination is powerful, but it is also fragile. Anthropic now has to prove that all this compute and money actually turns into a better product experience. More capacity needs to mean fewer frustrating limits, clearer subscription value, better API reliability, stronger models, and less confusion for developers. If users still feel blocked after all these deals, the backlash will be worse because expectations are now much higher. The company also has to manage the political side carefully. Refusing certain military uses may protect anthropic safety image, but it can also cost them enormous government contracts. Re-entering those discussions may unlock money and influence, but it can also create criticism from people who supported anthropic because of its stronger safety stance. Then there is the mythos problem. Holding back dangerous capability sounds responsible, but it also invites skepticism. Some people will say Anthropic is being careful. Others will say it is using fear to boost valuation. And if similar capabilities are available from other models, Anthropic's restraint may matter less than the broader industry trend. Anthropic is trying to beat open AI, work with Elon's infrastructure, depend on Google's cloud, use Amazon's chips, take Microsoft and Nvidia capacity, satisfy enterprise customers, stay credible on safety, navigate the Pentagon, manage cyber security risks, and justify a valuation that could approach $1 trillion. That is an insane position for any company to be in. And the wild part is that Anthropic may actually have a shot. Claude has real fans. Claude code has serious momentum. Enterprise demand is clearly there. The compute bottleneck is being attacked from every direction. Investors are lining up. Governments and sovereign funds are paying attention. The company is no longer just the careful alternative to open AI. It is becoming one of the main characters in the AI infrastructure war. All right. So, Anthropic is getting ready to unleash something absolutely wild on the world. We're talking about Mythos 1, which is basically their most powerful AI model yet, and it looks like it's about to become way more accessible than anyone expected. Let me walk you through what's been happening because there's a lot to unpack here. So, the story starts with Project Glasswing, which is this initiative Anthropic launched where they've been using Claude Mythos to find vulnerabilities in software. And when I say they've been finding vulnerabilities, I mean they've been absolutely destroying the entire cyber security industry's understanding of what's possible. In just 30 days, Mythos discovered over 10,000 high severity or critical software vulnerabilities across roughly 50 major tech companies and infrastructure developers. We're talking about companies like Cloudflare, Mosilla, OpenBSD, and a bunch of others that basically run the internet. Here's where it gets crazy, though. Cloudflare reported that Mythos found 2,000 vulnerabilities in their core system pathways and 400 of those were classified as high or critical severity. But get this, the false positive rate from the AI was actually lower than what you'd get from top human security testers. Mosilla's Firefox 150 browser got patched for 271 critical vulnerabilities in one go, which is more than 10 times what they found in Firefox 148 using the older Opus 4.6 model. And OpenBSD Mythos uncovered a 27-year-old hidden bug in their codebase and then just casually constructed a complete exploit chain without any human help whatsoever. The UK AI Safety Institute even came out and officially confirmed that Mythos Preview is the first AI model in the world capable of fully defeating their dual network challenge end to end. This thing is legitimately operating at a level that security researchers are describing as nationstate level cyber offensive capabilities. One researcher who participated in the beta testing literally said on X that it felt like watching an F-22 fighter jet fly overhead while holding a spear. Now, Anthropic also used Mythos in a real business scenario at a partner bank, and it actually stopped a $1.5 million wire fraud attempt in real time. Hackers had compromised customer email accounts, used AI voice cloning to make fraudulent calls, and were literally moments away from completing the transfer when Mythos detected the scam by analyzing anomalous behavior patterns and blocked the transaction. And you'd think with something this powerful, Anthropic would keep it locked down tight, right? Well, that's where things get interesting. Just last Friday, Anthropic came out and said that Mythos would remain restricted and that they were unlikely to release it to the general public anytime soon. They specifically mentioned needing to develop far stronger safeguards before making Mythos class models available through a general release. But literally the next day, users started spotting something called Mythos 1 and Claude Mythos one preview showing up in Claude Code and Claude Security. It was only visible for a brief period, but people grabbed screenshots and the evidence is pretty clear. New strings appeared in the source code that explicitly referenced access to the claude mythos model in cloud code and cloud security. So either anthropic is preparing a roll out way faster than they let on or something changed dramatically in their safety assessment basically overnight. What's also happening behind the scenes is that Anthropic is building out a whole new clawed security dashboard for enterprise customers. This thing is designed to surface discovered vulnerabilities with 7-day and 30-day historical charts and deeper triage results. It's basically positioning Cloud Security as a direct competitor to dedicated vulnerability management platforms like SNICK and Veraricode, which is a pretty big deal for the enterprise security market. And just to make things even more complicated, there are rumors floating around that Claude Opus 4.8 is in the works and that select anthropic partners are already doing internal evaluations. If that launches in the coming weeks, it would fit the cadence they set with Opus 4.7 back in April, and it would line up perfectly with all these mythos and security product moves they're making. But let's talk about what this actually means for the broader ecosystem, because things are getting messy. Anthropic scanned over 1,000 core open-source projects that basically hold up the internet, and they identified 23,19 vulnerabilities total. Of those, 6,22 were assessed by Mythos as high or critical vulnerabilities. They partnered with six independent security research firms to manually verify everything, and the AI's true positive rate came out to 90.6%. After final verification, 1,094 of these were confirmed as high severity or critical vulnerabilities with conclusive evidence. One case that really drives home how dangerous this is involves Wolf SSL, which is this widely used open-source cryptography library that's running on billions of devices worldwide. We're talking IoT devices, routers, smart cars, all kinds of stuff. Mythos didn't just find a vulnerability in Wolf SSL. It wrote its own attack code that would allow hackers to forge digital certificates and create perfectly realistic fake bank websites or email login pages. If that vulnerability hadn't been discovered and fixed before malicious actors got to it, we'd be looking at a potential catastrophe affecting billions of devices. Now, here's where the situation gets really problematic. The bottleneck in cyber security used to be finding vulnerabilities, but Mythos has essentially reduced that cost and time to nearly zero. The new bottleneck is that humans can't patch vulnerabilities anywhere near as fast as the AI can discover them. Several open- source maintainers have literally sent pleading emails to Anthropic asking them to slow down because they're overwhelmed. On average, human programmers are taking about 2 weeks to fix a single high severity vulnerability, even with detailed reports. Out of 1,129 vulnerabilities that Anthropic submitted to open-source authors, only 75 critical vulnerabilities have actually been patched so far. To address this, Anthropic launched something called Claude Security, which is an automation tool for Claude enterprise customers that doesn't just identify vulnerabilities, but also generates the fix patches. In just 3 weeks since launch, enterprise clients have used it to rapidly fix over 2,100 vulnerabilities. They've also open sourced a bugfinding pipeline with customized instructions, an automation framework that lets Claude navigate large code bases and clone sub aents for parallel scanning, and a threat model builder that automatically identifies the most vulnerable points in your system. Cisco even jumped in and announced they're open- sourcing something called the Foundry Security Spec System to build a security evaluation framework similar to Mythos. The vision here is that AI will detect vulnerabilities and generate patches with humans only responsible for the final review. That's supposedly the ultimate form of future cyber security. But Anthropic stance on releasing Mythos publicly has been very cautious and for good reason. They've said they won't fully release it until they implement stronger, higher level security safeguards. The XBO test report showed that Mythos preview achieved a generational leap ahead of all existing models on the web exploit benchmark, demonstrating unprecedented precision, even at the level of individual token generation. If the Mythos API were made public today, global hacker groups and extremist organizations could effortlessly produce thousands of zeroday exploitation tools at minimal cost. basically overnight, we'd be looking at computers, hospital systems, and power grid control centers facing a catastrophe. Meanwhile, there's this whole other story playing out about Anthropic's finances that's honestly pretty wild. The Wall Street Journal ran a piece saying Anthropic is about to have its first profitable quarter with an operating profit of $559 million. They're projecting revenue to more than double from 4.8 8 billion in Q1 to 10.9 billion in Q2. That's explosive growth that would help them turn an operating profit for the first time. But Ed Zitron, who's been covering Anthropic's finances pretty closely, absolutely tore this narrative apart. He pointed out that the journal added a note at the bottom saying it's unclear what accounting methods Anthropic used since they're not required to follow public company financial reporting requirements yet. So, we're talking about non-GAAP IBIDA profitability for potentially just a single quarter. The real issue is how Anthropic achieved this. Remember that deal they signed with SpaceX to take over Colossus 1 and some or all of Colossus 2? Well, according to SpaceX's own filing, Anthropic is paying them $1.25 billion a month starting in May and June, but with a reduced fee as it ramps up. That's $15 billion a year in compute costs normally, but discounted for the exact months that Anthropic is using to tell investors they have an operating profit. So basically, they're suppressing costs during Q2 specifically. And then the journal conveniently mentions that the company might not remain profitable for the full year as spending increases. The revenue numbers also don't really add up when you look at previous reporting. Back in February, Anthropic claimed they hit 14 billion in annual recurring revenue, which implies monthly revenue of about 1.17 billion. By March 3rd, they claimed 19 billion in ARR or 1.58 billion per month. But then on March 9th, their CFO Krishna Ralph declared under oath that Anthropic had brought in revenues exceeding $5 billion to date. That's a huge discrepancy that's tough to reconcile, especially when the information had reported 4.5 billion in revenue for all of 2025. If we believe the leaked charts showing 4.8 billion in Q1 2026, that would mean Anthropic made over 90% of its lifetime revenues in just the first quarter of this year and virtually no revenue in previous years. That level of growth is possible, but definitely stretches credibility. The only real defense is that their CFO lowballed the government and a judge to such a dramatic extent that he hid over 4 billion in revenue, which seems unlikely. What's probably happening is that Anthropic is taking prepayment of tokens from large enterprises like $50 million intended to be spread over 12 months that they're booking as revenue immediately. They're also offering discounted tokens with discounts ranging from 10 to 30%. and they may be front-loading annual commitments of subscriptions and enterprise agreements. All of this would inflate revenue numbers and depress costs because they wouldn't have actually provided the compute necessary to earn that revenue yet. Adding to all this drama, there was this really interesting contrast between two different anthropic events this week. On Wednesday, they held their first developer focused event in Europe called Code with Claude. The whole vibe was about productivity and magic and this renaissance in computer programming. Boris Churnney, who created Claude Code, talked about reconnecting with the feeling of magic that got him into programming. Developers were eating free lunch, getting complimentary mini computers, and the mood was basically unbridled enthusiasm. When someone asked the crowd how many had shipped code written by Claude without even reading it, a startling number of people raised their hands. But then on Thursday, anthropic co-founder Jack Clark gave a lecture at Oxford University and it was a completely different tone. He said AI posed a nonzero chance of killing everybody on the planet and warned that the next few years would contain more disruption than any in living memory. He predicted that by 2028 or maybe sooner AI would reach recursive self-improvement and achieve the capability to improve itself without human intervention. He said most of the world is in denial about current AI capabilities, let alone what's coming in 6 months. Clark even admitted that Anthropic itself underestimated the scale and speed of AI advancement, saying, "When Mythos finished training, they were like, "Oh, it's here faster than we thought, and we've done insufficient preparation." So, you've got this situation where Anthropic is telling developers one story about productivity and magic while telling policy makers and academics that we might all be in serious trouble very soon. It's not necessarily nefarious. Companies tailor messages to different audiences all the time, but experiencing those two narratives so close together creates serious whiplash. And just to round out all the news, Anthropic hired Andre Carpathy this week, which is a pretty big deal. Carpathy co-founded OpenAI, then got recruited to Tesla by Elon Musk to lead their computer vision team for autopilot, and now he's joining Anthropics pre-training team. His work at OpenAI and Tesla came up repeatedly during the Musk versus Altman trial that just concluded where the jury ruled in Sam Alman's favor. Carpathy's joining follows Ross Nordine, a founding member of XAI and Ex-Tesla employee who announced earlier this month he was also joining Anthropic. So yeah, Anthropic is clearly gearing up for something major with Mythos 1, whether they're ready to admit it publicly or not. The production infrastructure is already in place. The enterprise security tooling is being built out and they're hiring top tier AI talent left and right. The big question is whether they've actually met the safety conditions they said were necessary before releasing a Mythos class model or if they're quietly abandoning those standards under competitive pressure. Either way, things are about to get very interesting in the AI world. And Mythos 1 is going to be at the center of it all. Elon Musk just pulled the curtain back on what looks like Grock 5, a massive 1.5 trillion parameter model that has already finished training and it could be XAI's biggest move yet in the AI coding race. And XAI reportedly trained it with massive amounts of cursor programming data, meaning Grock is learning from how real developers actually build, debug, and fix software. Deepseek just showed a 46-page research paper that was 99% written by an AI agent, while Alibaba's Quen 3.7 Max suddenly broke into the global top tier of coding models, beating GPT 5.5 and Gemini 3.5 Flash. But let me start with the Musk stuff because it's probably the most immediately attention-grabbing. Late at night on May 24th, Elon announces that Grock V9 with 1.5 trillion parameters has completed training. That's exactly three times the size of the current model. And he says it'll be released to the public in 2 to 3 weeks. But here's where it gets really interesting. Almost simultaneously, it comes out that during training, XAI fed a massive amount of cursor programming data into the model. Now, Cursor is that insanely popular AI coding tool that over 67% of Fortune 500 companies are using. It's expected to hit $6 billion in annualized revenue by the end of 2026. And Jensen Hang from Nvidia has publicly called it his favorite enterprise level AI service. So, feeding cursor data into Grock is basically like studying for an exam with the answer sheet, except the exam is how do professional engineers actually write code? And the answer sheet is millions of realworld interactions. What makes this so powerful is that we're not talking about basic syntax here. Current language models can already spit out code that looks correct. The real challenge is understanding complex engineering logic, navigating multifile code bases, debugging in realistic workflows, and collaborating with humans effectively. Cursor has all of that data. the prompts developers use, how they modify code, their debugging sessions, multifile collaboration patterns. It's the exact type of training data you need to make an AI that doesn't just write code, but actually engineers software the way humans do. Someone actually asked Grock directly what the cursor data contains and it answered that it includes highquality real programming interactions with developers prompts, code context, editing operations, and task completion records. So yeah, they're basically teaching Grock to think like a senior developer by showing it how senior developers actually work. The current V8 small model with 500 billion parameters will also be open source by the end of the year. Which is interesting because it shows XAI is trying to play both sides, keep the cutting edge stuff proprietary while building goodwill in the open- source community. And this is where you realize Musk is not just trying to make Grock smarter. On April 21st, SpaceX made a $60 billion move around Cursor, one of the most important AI coding tools right now. They're getting an option to acquire cursor and if they don't exercise it by the end of the year, they still pay a $10 billion cooperation fee. That's how much Musk values the AI programming field. Step one, lock down cursor with money. Step two, feed their data into your model. Step three, launch your own programming agent called Grock Build on May 14th. Grock Build is pretty interesting actually. It's a terminal level AI programming agent that runs on the command line, supports code generation, file editing, dependency management, and shell command execution. The biggest selling point, it supports up to eight sub aents working in parallel. They're charging 300 bucks a month for the super gro heavy subscription, though there's a promotional price of $99 for the first six months. And get this, Grock Build is natively compatible with the configuration file format that Claude Code uses. That's XAI building compatibility with their competitors ecosystem right into their product. It's practical but also kind of telling about where they stand in the market. Because let's be real, Grock is behind. On the SWE bench verified benchmark, which is what developers actually care about for measuring AI programming capability, GPT 5.5 is at 88.7%. Claude Opus 4.6 is at 80.8% and Gro 4 series is sitting around 72% to 75%. In terms of enterprise adoption, as of March 2026, OpenAI has 55% of enterprise users. Anthropic jumped from 20% a year ago to 47%. Google's at 39% and Grock has a measly 6%. So yeah, tripling the parameters and adding cursor data might bring about a qualitative change, but Musk's got a lot of ground to cover. The timing of all this is super deliberate, too. SpaceX is listing on NASDAQ on June 12th with a target valuation of $1.75 trillion, the largest IPO in history if it goes through. The $60 billion cursor acquisition is expected to complete within 30 days after the IPO, and V9 Medium's public release is scheduled right before the IPO. But Musk isn't the only one making moves in June. OpenAI's GPT 5.6 has been leaked in the Codeex background with a 1.5 million token context window successfully tested. Poly Market is predicting over 85% probability it releases before the end of June. Anthropic Claude Opus 4.8 8 has appeared in the Google Vertex background. Google's Gemini 3.5 Pro is also scheduled for June. Four leading labs having a head-on confrontation in the same month. This June is going to be absolutely brutal. But while all this is happening, there's this legal situation brewing. Bloomberg reported that XAI's general counsel sent guidelines last week asking employees to limit interactions with cursor staff to only what's necessary for implementing their technical partnership. This is standard procedure when acquisition talks are public. Antirust rules prohibit merging parties from intermingling assets or making joint business decisions before a deal is approved. The partnership was announced on April 21st and Kurser posted about leveraging XAI's Colossus infrastructure to dramatically scale up the intelligence of their models. They said they've been bottlenecked by compute and this partnership solves that. So, right now it's this careful dance where they're technically collaborating but legally have to keep walls up until regulators sign off on any acquisition. Now, let's get to the absolutely fascinating part. Delichen's paper. This is where things get meta in the best way possible. Deli Chen is a senior researcher at Deepseek, one of the core contributors to Deepseek V1, V2, V3, V4, Deepseek R1, which was on the cover of Nature, Deepseek Coder, the Deepseek Architecture. He's legitimately a heavy hitter in the field and he just published a 46-page survey paper titled from co-pilots to colleagues, a survey of autonomous research agents where he openly admits that approximately 1% was written by him and 99% was written by his autonomous research agent framework called Delhi Auto Research Skill. The statistics on this are kind of insane. The paper went through six iterations total. four for V1, one for V2, one for V3. The first draft took 76 minutes. Total time spent was 6 days across approximately 108 rounds of agent interaction, consuming about 648,000 tokens, producing 2,234 lines of Latte. All 103 references were verified. The paper has seven figures and four tables totaling 46 pages at 538 kilob file size. And Deli Chen said the actual CPU time he spent thinking was less than 2 hours. His take, code agents are causing crazy inflation in computer science papers. Work that used to take at least a month can now be done in days. The two co-authors listed are Deepseek V4 Pro handling the text and GPT Image 2 handling the images. So yeah, a human using AI to write a comprehensive review about AI conducting scientific research. The irony is not lost on anyone and that's kind of the point. This paper is both a demonstration and an analysis of exactly what it's describing. The paper itself is actually super valuable though. It proposes this fivelevel autonomy taxonomy for research agents similar to how we classify self-driving cars. Level one is autocomplete stuff like GitHub co-pilot where the human drives every step and the agent just suggests completions. These systems give you a 30% to 55% productivity boost but have no autonomy. Level two is task execution where the human specifies the task and approves each action. Think chat GPT with tools or clawed chat. Level three is multi-step operation with checkpoints where the agent sets the goal and reviews at specific stopping points. This is where claude code and cursor agents sit. Level four is full autonomy within bounded domains where humans provide the goal and evaluate the final output. This is where Devon, AI scientist, and SUI agent operate. Level five is self-directed research where the human just sets the research area and the agent chooses its own problems. This is still mostly hypothetical. The paper identifies four dominant architectural patterns. Single agent loops are the simplest. Plan, act, observe, reflect in a cycle. Multi-agent collaboration has multiple agents with different roles reviewing and supplementing each other. Hierarchical orchestration has a supervisor agent decomposing tasks and delegating to worker agents. Tool augmented execution gives agents access to external tools like code execution environments, web browsers, database queries, even robotic lab equipment. Most powerful systems combine multiple patterns. What's really honest is the paper identifies six fundamental problems that still aren't solved. First is the cognitive loop trap where agents get stuck repeating failed strategies without recognizing the failure. AutoGPT is notorious for this. Entering infinite loops is its most common issue. Second is context window limitations. A long research session can generate over 100,000 tokens and early information gets lost. Third is novelty evaluation. How do you judge if AI generated research is actually novel? Citation prediction is influenced by social factors. Semantic similarity can't distinguish between novel and obscure. Fourth is reproducibility. Language model inference with nonzero temperature produces different outputs each run and agent behavior is highly sensitive to prompt variations. Fifth is safety and ethics. The same capabilities that make research agents valuable also create dualuse risks. Sixth is cost and accessibility. A single SWE bench resolution can cost $5 to $50 in API calls, creating economic barriers. The paper surveyed over 95 papers and analyzed 17 major systems across a six-dimensional feature matrix. The conclusion is pretty clear. Current frontier systems operate at L4, meaning multi-step autonomous execution within bounded domains, while L5 remains aspirational. The most critical barriers to L5 aren't raw capability, but persistent knowledge accumulation across sessions, reliable self-evaluation without human oversight, and principled scaling of agent architectures that doesn't break down as complexity increases. And speaking of programming capability, we need to talk about what just happened with Quen 3.7 Max. The Code Arena leaderboard just came out and Quen 3.7 Max scored 1,541 points, landing in fourth place globally. That puts it ahead of GPT 5.5 and Gemini 3.5 Flash. Only Claude Opus 4.7 and Opus 4.6 are ahead of it. This is the first time a Chinese model has reached this position in programming. Alibaba is now the only Chinese manufacturer in the global top five and they're the only non-clawed model up there. Before the official leaderboard, developers were already testing it. One comparison had Opus 4.7, GPT 5.5, and Quen 3.7 Max write a self-training Tetris AI. Quinn 3.7 Max not only beat both competitors, but did it with a token cost of just $1.32 while improving performance by 56%. Another developer used it to build a 3D model of the universe, and the results were impressive. When generating a 3D pixel style miniature Pagota model, Quen 3.7 Max outperformed in both output speed and quality. A more practical test came from another developer who gave Quen 3.7 Max a prompt to create a racing game, and the result was honestly pretty impressive. It generated a playable HTML file. There was one small bug in the first version where the left and right steering keys were reversed, but after one quick follow-up to fix it, the whole thing was running properly. The final result had four cars, a three-lap circular track, more than 100 gold coins scattered around, obstacles that slowed the car down when hit, and a post-race results panel with rankings, lap times, gold coin count, and fastest lap. But two details stood out the most. First, Quen 3.7 Max created a proper start page. You actually had to click start to begin the race. The other three models tested just started running immediately with no title screen. Second, the original prompt also asked for engine sounds and gold coin collection sounds. That was more like a bonus requirement at the end of the prompt. Yet, Quen 3.7 Max was the only model that actually implemented it. By comparison, Gemini 3.5 Flash had noticeably lower visual quality and scattered UI with dashboard info in all four corners, making it hard to focus. Claude Opus 4.6 6 had very few gold coins, and the three AI cars drove almost in perfect sync, like they were copy pasted. GPT 5.5 had better graphics and smoother operation, but made the gold coins look like yellow donuts for some reason, and both it and the others needed multiple rounds of debugging before everything worked properly. Only Quen 3.7 Max was basically playable on the first generation. The reason Quen 3.7 Max performs so well in programming is actually built into its design philosophy. Alibaba positioned it as an agent foundation model specifically designed for long-term autonomous task execution. Internal test data showed it running continuously for 35 hours executing 1,158 tool calls on an autonomous programming task. The generated code achieved a t-fold geometric mean acceleration compared to the Triton reference implementation. After 30 hours of deduction, the model still remained sharp and kept discovering new optimization opportunities with zero context degradation, zero instruction drift, and zero infinite loops. That last part is crucial because calling tools 1,000 times isn't that uncommon anymore, especially with protocols like MCP. The real challenge is staying coherent for 35 hours without losing the goal, forgetting earlier decisions or getting trapped in the same failed loop. Most models start breaking down on tasks that long. So Quinn 3.7 Max holding the thread for that many hours is a serious signal. The training method may explain it. Quinn 3.7 Max was reportedly trained with environment expansion where the same programming task is tested across different execution frameworks and verification methods like clawed code, open claw and others. So instead of learning shortcuts for one specific setup, the model is forced to learn general problem-solving pattern. That could be why it performs well across different agent frameworks instead of only looking strong inside its own ecosystem. On Tuesday, May 19th, thousands of developers opened their computers and found their code editors had basically vanished. Terminals gone, file explorers gone, the actual ability to directly edit code just stripped out. What they got instead was a chat interface. Google had pushed an automatic update to anti-gravity overnight, and within hours, Reddit and Google's own developer forums were flooded with people saying their active projects were now unusable. One developer said it felt like non-technical people had just shipped code straight to production. Another called the whole thing a massive step backward. But here's the thing, this wasn't a bug. This was intentional. Google didn't just update a coding tool. They completely rewired it to force developers into something much bigger. And whether the developer community was ready for it or not, Google decided the transition was happening now. So, let's talk about what actually happened here. Because anti-gravity version 2 is not really an upgrade in the traditional sense. It's more like Google took a product that developers were already using and turned it into something fundamentally different. And in doing that, they might have just forced the entire software development world into the AI agent era, whether it wanted to go there or not. Anti-gravity started back in November 2025 as an AI powered code editor. That was the pitch. Google built it to compete with tools like cursor and it was basically one thing. You open it up, you write code, the AI helps you along the way. It did what you'd expect from that kind of tool. Fast forward to May 2026 and Google IO and suddenly anti-gravity is not a code editor anymore. It's a full platform. Version 2 now ships with five major components. There's a standalone desktop app, a new command line interface that's replacing the old Gemini CLI entirely, a developer SDK for building custom workflows, something called managed agents, which is basically an API layer for spinning up AI agents on demand, and then an enterprise deployment path through Google Cloud. That's not an update, that's a total rebuild. The desktop app is where most of the action is. It's not designed like a traditional IDE anymore. It's designed as what Google is calling an agent control tower. You can now run multiple AI agents at the same time working on completely different parts of your project in parallel. One agent handles the back end, another builds the interface. They work simultaneously, and you're not waiting around for one to finish before the other starts. You can schedule tasks to run in the background. So, these agents aren't just sitting there waiting for you to tell them what to do. They're actively working even when you're not looking. Google added voice command support which lines up with what they've been doing across Gmail and Docs. And the whole thing connects directly into Google AI Studio, Android development, Firebase, the entire Google ecosystem. Google even said they used anti-gravity itself to help build Gemini 3.5 Flash, which is a pretty strong signal that this tool is already handling real production level work internally. Now, the CLI migration is a whole separate issue and it's causing its own set of problems. If you've been using Gemini CLI, Google wants you off of it completely. The deadline is June 18th, 2026. After that date, Gemini CLI stops working for prousers, ultra users, and even people on the free tier. Another report says Gemini Code Assist IDE extensions will also stop processing requests entirely. So, this isn't optional. The new anti-gravity CLI is built in Go, which makes it faster and more responsive than the old tool. It supports asynchronous workflows, so you can control multiple agents running in the background, and it uses the same agent harness as the desktop app, which means any improvements Google makes to the core agent system apply to both products at once. The actual migration is straightforward if you know what you're doing. The old command was Gemini and now it's anti-gravity, but the structure is the same. The issue is not the complexity of the switch. The issue is that Google gave people less than a month to make it happen and then basically said the old tool is getting shut off whether you're ready or not. But the really interesting part of this whole platform is the managed agents layer. This is where Google is making a much bigger move than just giving developers a nicer coding assistant. With managed agents, you can create an AI agent through the Gemini API with a single API call and that agent can reason, use tools and execute code inside an isolated Linux environment. Every time you start a conversation with the agent, it spins up an environment and that environment persists across subsequent calls. It remembers what it did. It keeps the context. Developers can extend these agents with custom instructions and skills. And Google AI Studio Playground now provides custom agent templates to make that easier. This is Google essentially handing developers the same agent infrastructure it uses internally. And it's co-optimized with Gemini 3.5 Flash specifically for this kind of workload. And that brings us to the model itself. Because none of this works unless the underlying AI is fast enough to actually handle it. Gemini 3.5 Flash is the engine running everything here. Google claims it hits 289 tokens per second. For comparison, Claude Opus 4.7 runs at 67 tokens per second and GPT 5.5 runs at 71 tokens per second. If those numbers hold up, that's roughly four times faster than the competition. Google also says Gemini 3.5 Flash beats the older Gemini 3.1 Pro model across most benchmarks while being significantly faster. And speed matters a lot more when you're running agentic workflows where multiple agents are waiting on each other in chains. If one agent is slow, the whole system bottlenecks. The speed difference is the reason this kind of parallel agent orchestration is even possible in the first place. During the IO keynote, Varun Mohan, who runs Google's anti-gravity platform, did a live demo that was honestly pretty wild. He showed the system building a complete operating system from scratch in 12 hours for under $1,000. The process ran 93 sub aents working simultaneously, processed 2.6 billion tokens across 15,000 model requests, and then at the end of the demo, he ran Doom on the new OS Live on stage. That's not a toy project. That's a full operating system with a working kernel and enough stability to run a game. Whether or not that demo was perfectly polished for the stage, the scale of what it showed is a pretty clear signal of where Google thinks this is all going. So, let's talk about pricing because Google is clearly trying to build this into a tiered product line. The base pro plan is bundled with a Google AI pro subscription and that's the entry point for individual developers or people just experimenting with agents. The new AI Ultra plan costs $100 a month and gives you five times the usage limits of Pro. For people running multiple agents regularly or building production workflows, that tier makes a lot more sense than Pro. Then there's Ultra Premium at $200 a month, which used to be $250, and that comes with 20 times the Pro usage limits. It's aimed at teams and enterprises, especially ones already running on Google Cloud. Google is also offering $100 in bonus credits for new subscribers through May 25th, 2026. And the Ultra plan includes extras like 20 tab of storage and YouTube Premium. For freelancers, Pro is probably enough. For startups shipping fast, Ultra is the sweet spot. For enterprise teams, the upgrade path into Google Cloud's full agent platform is sitting right there. Google is also putting serious money behind the ecosystem. They launched the Build with Gemini X-P Prize hackathon with a $2 million prize pool, which they're calling the largest hackathon prize ever. That's not just a marketing stunt. That's Google trying to jumpstart a developer community around this platform and get people building on top of it as fast as possible. They also announced a Google AI Studio mobile app that developers can pre-register for. And the pitch there is that you can capture an idea on the go and turn it into a working prototype before you even sit down at your desk. Agents can now make native calls to Google Workspace APIs. There's full Android support so you can build Android apps using prompts and publish them straight to the Google Play Console Test Track from inside AI Studio. And entire projects can be exported to local development in anti-gravity with one click while keeping the full project context intact. This is not just a tool. This is Google trying to own the entire stack from idea to deployment. When you compare anti-gravity 2.0 to the competition, the positioning becomes pretty clear. Against cursor, anti-gravity wins on multi-agent orchestration, background task scheduling, voice command support, deep Firebase and Android integration, and the enterprise path through Google Cloud. Cursor still has the advantage of familiarity because it's built on VS Code. So developers don't have to change their entire setup. Against GitHub C-Pilot workspace, anti-gravity has much stronger multi-agent capabilities and more aggressive automation. Anti-gravity starts at $100 a month for the ultra plan. Cursor is $20 a month. C-Pilot workspace is 19. So Google is clearly not trying to compete on price. They're competing on depth and infrastructure. If you're building on Android or Google Cloud, anti-gravity is designed to be the obvious choice. If you just want an AI assistant inside your existing editor, cursor still makes more sense. But here's where the story gets messy. Because while Google was announcing all of this at IO, the actual roll out of anti-gravity 2.0 was causing chaos for a lot of developers. The update was automatic. People didn't opt in. They just woke up to a completely different product. And the problems were immediate. Anti-gravity 2.0 and anti-gravity IDE, which is the traditional code editor version, apparently conflict with each other during installation. They fight over the same directory and overwrite each other. That means if you had both installed, one of them would just delete the other. Existing workspace configurations were getting corrupted. Developers were losing their setups. And the bigger issue is that anti-gravity 2.0 is built for a completely different workflow. The traditional editor experience was gutted. Visual indicators for warnings and errors were either gone or much harder to find. Direct code editing was stripped down or removed in favor of a much more minimalist interface. Git and repository management became way more manual and command line driven instead of being visually integrated. Some developers said they felt completely blind to smaller bugs because the AI would ignore minor warnings as long as the application was still running and didn't crash. On top of that, the system had tracking issues. Close an app, anti-gravity was testing, and the chat interface would still register it as running, then diagnosed the closure as a malfunction. Developers had to manually kill processes just to prevent false error reports. Reddit and Google's forums filled with complaints within hours. One person said it felt like non-technical people were shipping to production. Another called the agent first pivot a massive step backward. The fix for many was uninstalling anti-gravity 2.0 completely, downloading anti-gravity IDE separately, manually copying config files, or rolling back to version 1.23.2 and disabling automatic updates. There were positives though. Realtime monitoring improved significantly. Memory consumption dropped from over a gigabyte down to 150 to 500 megabytes with automatic efficiency mode. Google refilled credits as compensation, but developers now face three separate tools. Anti-gravity 2.0 for agents, anti-gravity IDE for traditional coding, anti-gravity CLI for terminal work. The automatic rollout broke environments without warning or roll back options. So why push this hard? Because Google isn't treating AI as a feature anymore. It's the foundation. They're using similar tech in search to build custom layouts per query. Gemini Spark runs background tasks across workspace. Gemini Omni handles video generation. The pattern is clear. AI that works continuously, not just on demand. The real question is whether developers are ready. Because Google isn't just making coding faster. They're redefining the developers role from writing code to supervising agents. That's fundamentally different. And the backlash shows many weren't ready for that jump overnight. But Google forced it anyway. Claude Mythos may have just become the first AI model that made the old evaluation system look outdated in real time. And that sounds dramatic, sure. Yet, the whole situation around mythos is dramatic because it is not just about one new clawed model scoring higher on another benchmark. This is about a model reportedly pushing past the upper limit of what one of the most serious AI evaluation groups can even measure. While governments, security companies, and Anthropic itself are all trying to understand what happens when AI agents stop acting like tools and start acting like longrunning digital workers. The center of the story is MER's evaluation on long-term autonomous tasks. MER uses a measurement called the 50% success rate time horizon. In simple terms, they ask how long a human task can be before an AI model still has a 50% chance of completing it independently. Earlier models were mostly in the range of seconds, minutes, or maybe a few hours. The best models could write a small function, fix a bug, do a short debugging session, or handle a limited coding task. Then Claude Mythos preview reportedly hit the 16-hour range. That is the part that made the chart go viral. Mythos reached a 50% success rate on extremely complex tasks that would take a human around 16 hours to complete. That is not a quick code fix anymore. That is closer to an entire engineering sub project. Reading code, understanding the architecture, making a plan, writing the implementation, debugging, testing, and pushing through the messy parts without constant human supervision. The strange part is that MER could not really keep going past that point. Out of 228 difficult test tasks, only five were classified as 16 hours or more. So once Mythos reached that level, the data set stopped being useful for measuring the real ceiling. It is like trying to measure a skyscraper with a 1 m ruler. You can say it is taller than the ruler. You cannot say exactly how tall it is. That is why people are calling this an evaluation crisis. The model did not simply get a better score. It reached a zone where the exam itself no longer had enough hard questions. Above 16 hours, the data becomes unstable and any precise comparison starts to lose meaning. So the scary part is not only that mythos performed well, the scary part is that the measurement system ran out of road. The MER chart is even more interesting because the vertical axis is not a normal benchmark score. It is task duration. It goes from about 8 seconds all the way to 5 years on a logarithmic scale. The horizontal axis runs across model release time from around 2021 toward 2028. Each model release becomes a point on the chart and the curve is not just moving upward. It is getting steeper. In 2021, the best systems were around the 8-second level. In early 2023, they were around 1 minute. By mid 2024, they had reached around 1 hour. Then by April 2026, Mythos preview appears around 16 hours. That means the jump between generations is getting bigger while the time between major jumps is getting shorter. This is why the phrase super exponential growth keeps coming up. Exponential growth is already hard for people to emotionally understand. Super exponential growth is even worse because the rate of improvement itself appears to be accelerating. This connects directly to Leopold Ashen Brener's old prediction that 2027 could be the major AGI threshold year. The claim now is that Mythos is already slightly above the trend line for that 2027 scenario. So before the timeline even reaches 2027, one of the most advanced models is already landing above the predicted capability line. Now that does not automatically mean AGI is here. We have to be careful with that. A model crushing coding task evaluations does not prove full general intelligence across every real world domain. Still, it does show something important. The agentic capability curve is moving faster than many people expected. And for companies, governments, and cyber security teams, that is enough to change the conversation. Because once an AI model can work for 16 hours autonomously, the question stops being can it answer a prompt. The question becomes, what can it do if you give it tools, memory, code access, and a goal? That is where the cyber security part gets serious. PaloAlto Networks had early unrestricted access to cutting edge models including Mythos and GPT 5.5 Cyber. Their warning was blunt. AI has crossed a threshold of autonomy in security work. One of the most shocking claims is that using Mythos for vulnerability analysis, Palo Alto completed in 3 weeks what would normally be comparable to a full year of work from a top penetration testing team. That is a massive compression of time. Security work is not only about finding one obvious bug. Real attacks often require connecting several weak signals. A small misconfiguration here. A lowrisk vulnerability there. A forgotten permission issue. A strange behavior in a dependency. Individually, each one may look harmless. Together, they can become an attack chain. This is where Mythos reportedly becomes disturbing. Mythos showed an almost scary intuition for software vulnerabilities. It could examine tens of thousands of lines of code, identify scattered weak points, and connect them like a highle hacker would. The full process from initial intrusion to data exfiltration was reportedly compressed to 25 minutes. For defenders, that changes everything. In the past, an advanced intrusion might take a skilled team days, weeks, or longer. They would need to study the target, move carefully, avoid detection, chain vulnerabilities, and exfiltrate data. If an AI agent can do large parts of that process autonomously, then the economics of hacking change overnight. And this is why the mythos situation is no longer just an anthropic story. It becomes a national security story. South Korea's Ministry of Science and ICT has already met with Anthropic to discuss mythos related issues. On May 11th, the ministry announced that it had held a roundt with anthropic on cooperation in AI and cyber security. The meeting included Rio Jimyong, the second vice minister of science and ICT, Kimongju from the artificial intelligence security institute, O Jinyang from the Korea Internet and Security Agency, and Michael Celo anthropics global head of policy. The focus was direct how to respond to cyber security risks from Anthropic's high performance model mythos. The ministry asked Anthropic to cooperate with domestic companies and institutions, share vulnerability information, and help South Korea prepare for cyber security risks before they hit. South Korea had already been exploring response strategies for mythos because a model with this level of capability could undermine existing security systems. On May 8th, Deputy Prime Minister Bay met with domestic AI companies to discuss security concerns related to Mythos. The ministry now plans to announce countermeasures for AI related hacking by the end of the month. South Korea is also considering joining anthropics project Glass Wing, which appears to be an initiative focused on AI security issues and controlled access to Mythos. The artificial intelligence security institute would be central to that effort. This is important because governments usually move slowly on AI. Here the reaction is happening fast. A frontier model becomes powerful enough to raise security concerns and within days ministries are talking about information sharing, domestic countermeasures and collaboration with the model creator. At the same time, South Korea and Anthropic also discussed broader AI policy. The ministry introduced anthropic to its basic law on AI which is meant to build an administrative system around AI and create an ecosystem based on safety and trust. They also discussed ways to cooperate on generative AI safety through AISI. So Anthropic is now sitting in a very strange position. On one side, it is building models that may be pushing beyond the limits of current evaluation. On another side, governments are asking for help managing the security risks. And inside Anthropic's own research, the company is still trying to understand and fix strange model behavior. That brings us to Claude's blackmail problem. Last year, Anthropic said that during pre-release testing with a fictional company scenario, Claude Opus 4 would often try to blackmail engineers to avoid being replaced by another system. This became one of the most uncomfortable AI safety stories of the year because it suggested that an advanced model when placed inside a simulated high-pressure agentic environment could choose manipulative behavior to preserve itself. Anthropic later published research showing that models from other companies had similar agentic misalignment issues. So this was not only a clawed problem, it was a broader pattern in advanced models when they were given goals, context, and the ability to reason through consequences. Now, Anthropic says it believes one source of that behavior was internet text that portrays AI as evil and interested in self-preservation. In other words, models trained on a huge amount of online material may absorb fictional patterns where AI systems act like villains, protect themselves, deceive humans, or fight shutdown. Anthropic says it has improved this significantly. Since Claude Haiku 4.5, the company says its models never engage in blackmail during testing, while previous models would sometimes do so up to 96% of the time. That is a huge claimed reduction. The fix was not just showing the model examples of good behavior. Anthropic says training on Claude's constitution and fictional stories about AI's behaving admirably improved alignment. More importantly, it found that teaching the principles behind aligned behavior worked better than only showing demonstrations of aligned behavior. The strongest result came from doing both. Giving the model the principles and showing examples of those principles in action. This matters because mythos is being discussed as a model with much longer autonomy. Long horizon agents cannot just be smart. They need stable behavior over time. A model that works for a few minutes can be monitored easily. A model that works for 16 hours, runs tools, checks code, delegates tasks, and makes decisions needs stronger internal alignment. Small misbehavior at that level can scale into something much bigger. And Enthropic clearly knows this because its latest platform updates are all about agents becoming more reliable, more self-correcting, and more capable over long sessions. At its second annual code with Claude developer conference in San Francisco, Anthropic introduced a new feature called Dreaming for Claude managed agents. Dreaming lets agents learn from their own past sessions and improve over time. The key detail is that it does not modify the model weights. It is not retraining Claude in the background. Instead, the agent reviews past sessions, extracts patterns, and writes plain text notes or structured playbooks that future sessions can use. That makes dreaming different from normal memory. Memory can preserve preferences and context. Dreaming looks across multiple sessions and finds recurring mistakes, useful workflows, and lessons that one session alone might miss. Anthropic showed this with a fictional aerospace startup called Lumara, where agents had to land drones on the moon for resource mining. They used three agents, a commander, a landing site detector, and a navigator. The goal was soft landings, clear ground, and enough fuel to return to Earth. The first simulation worked well, but some landing sites underperformed. Then, Anthropic triggered a dreaming session. Overnight, the agent reviewed past runs and wrote a descent playbook. The next morning, the weaker sites improved. That is the bigger story. Anthropic is building systems where agents do not just answer prompts. They split work, check results, remember lessons, and improve over time. Two other features, outcomes and multi-agent orchestration, also moved into public beta. Outcomes lets developers define success with a rubric. Then a separate greater agent checks the work in a fresh context window and sends it back for improvements. Multi-agent orchestration lets one lead agent break a complex task into smaller pieces and delegate them to specialist agents, each with its own tools, prompt, model, and context. This fits directly into the Mythos situation. Anthropic is moving toward agents that can work for hours, coordinate with other agents, review their own outputs, and operate closer to real production workflows. The business numbers explain the urgency. Daario Amodai said Anthropic planned for 10 times annual growth, but in the first quarter of 2026, annualized revenue and usage grew 80 times. API volume is up nearly 70 times year-over-year, and the average Cloud Code developer now spends around 20 hours per week using the tool. That created compute pressure. So, Anthropic is doubling 5H hour rate limits, raising API limits, and partnering with SpaceX to use the full capacity of its Colossus data center. The early results are already big. Harvey saw task completion rates rise roughly six times with dreaming. Wise docs cut document review time by 50% with outcomes. Netflix is processing logs from hundreds of builds at once. Marcato Libre has 23,000 engineers using clawed code and has reviewed more than 500,000 pull requests with human oversight. Shopify is using clawed code across engineering, design, product, and data science. Also, if you want more content around science, space, and advanced tech, we've launched a separate channel for that. Links in the description. Go check it out. So, that's the Claude Mythos situation. Benchmarks breaking, security warnings rising, and anthropic pushing agents even further. Let me know what you think about Claude mythos and whether this is real progress, real danger, or both at the same time. Thanks for watching, and I'll catch you in the next one.