Opus 4.8 has just been released. Here's how to use it!

If you don't change how you use Claude Opus 4.8, you'll burn through your data allowance in minutes. So, how do you use Claude Opus 4.8 effectively? That's exactly what we're going to talk about in this video. There are two sides to the same coin. There's the marketing side, where Anthropic presented the release of Claude Opus 4.8. But there's the hidden side, the one no one will read, but I've done it for you. In this video, I'm going to tell you the truth, what Anthropic hid from you about Claude Opus 4.7, but especially about Opus 4.8: the arrival of dynamic workflows, the new "Ultra Code" feature, the one that will allow companies, developers, and engineers to stop working and use Claude Opus 4.8 to do their jobs. That's what Anthropic promises with Claude Opus 4.8. Are we getting close to an AGI system? A system capable of reasoning, working, and thinking for us? That's more or less how they introduced us to this new model, Claude Opus 4.8. But let me tell you the truth: you have to read between the lines, especially what Anthropic doesn't show in the documentation. Now, I'm not saying there aren't any improvements, I'm saying there are things to consider because otherwise, using Claude Opus 4.8 will be very expensive. The prompting has changed, the model's behavior has changed, and if you don't make these adjustments, you're headed for disaster. Regarding prompting, Anthropic explains that for Claude Opus 4.8, there's a relationship between the reasoning system and verbosity. This means that if you set the reasoning level to maximum, you'll get a more in-depth but also longer answer. If you set the effort to "medium," you'll get less thorough reasoning and shorter answers. But beyond that, Claude Opus 4.8's reasoning effort calibration system is currently linked to the model's ability to trigger tool management. This means that if you set Claude Opus 4.8 to "low" or "medium," it won't trigger any tools. It will automatically use its knowledge, that of the training data. And in this case, its knowledge cutoff, or knowledge base, is January 2026. Its token output window is 128,000 tokens, the same as Claude Opus 4.7. Its context window is 1 million tokens. So here, I'm going to reveal a first flaw, a first flaw that leads me to believe Anthropic is hiding something about Claude Opus 4.8's capabilities. The price is already the same as Claude Opus 4.7, but Anthropic pulled the same stunt with the launch of Claude Opus 4.7. I'm talking about the MR VRC2. So, if that doesn't mean much to you, imagine the AI's ability to handle a very large context of 1 million. Feel free to boost, like, and share this video. I'm counting on you. A like, a thoughtful comment—I'm always happy to reply. And for those who want to transform the way they learn to use artificial intelligence, because understand this: AI isn't going to replace you; it's someone who knows how to use it and who will take your job. For a very long time, we said that a degree protects and creates jobs. That's no longer true with AI. This is changing in two ways. First, AI is completely revolutionizing work today. Just look at how much Mythos, that is, Anthropic's super AI in the United States, is scaring everyone. That's the first point. So, AI is revolutionizing everything and changing the current paradigm regarding labor laws. Second, unemployment is rising again. Unemployment is rising, and because of AI, young people are no longer being hired. We don't talk about it, but the problem of youth unemployment is becoming one of the top priorities in France. I have children, I have a son who is 20 years old, and I worry about him a lot because of it. Regarding career prospects, despite a good degree, for a very long time, it was said that a degree protected and created jobs. This has become false with AI, because companies have made a choice: to save money. Why? Because the chosen policy is that employing someone is too expensive. Consequently, they will seek a specific profile: someone who knows how to configure AI agent systems. So, do you understand? Forget basic prompts. Forget "do my work for me, you're an expert." Today, the result is being able to go from beginner to pro in 15 days with 80 hours of training at your own pace. The latest AI agent systems are available on Claude, ChatGPT, Codex, Claude CLI, Gemini—in short, the best of AI. In your hands, you hold the future, and the future belongs to those who understand the impact, the revolution that artificial intelligence is having on the job market. Through this ecosystem, I will help you prepare for the Claude Code 101 exam and the Gemini Prompt Engineering exam. These are two qualifications you can add to your CV to stand out in the job market. All updates are included. All the information is in the description. I'm counting on you. A like, a boost, and we'll continue the video. This issue of youth unemployment is a national scandal that not everyone is aware of. If we look at Switzerland: two and a half times more industrial GDP, a trade surplus, and a youth unemployment rate of 3.2%. In France, it's now at 21%. Youth employment rates in France: 35%, in Germany: 51%, in the Netherlands: 75%. Now, they overwhelmingly want to start their own businesses, regardless of their level of education. AI is capable of handling very large contexts if it knows how to retrieve information, and that's what ChatGPT 5.5 demonstrated. It was able to retrieve 74% of the information in a context of one million tokens, while Claude Opus 4.7 only retrieved 32%, whereas it had previously been able to do 80%. And this was never disclosed to us, but the problem has just resurfaced in exactly the same way with Claude Opus 4.8. So, I'm not saying everything is bad, but I am saying that this information is missing, and consequently, it undermines the entire marketing pitch about the million tokens in the window. Because this whole architectural logic rests on two elements. Here they are. Anthropic, with Claude Opus 4.8, has just launched the dynamic workflow in Claude Code. Imagine an AI capable of working for 10 days straight, where Claude independently generates orchestration scripts to support dozens—up to 100 subordinate agents— working in parallel during a single session. These agents can collaborate, communicate with each other, and handle massive volumes of code, up to nearly 1 million lines and 500 files, thanks to the new "Ultra Code" feature. This dynamic workflow, with its more precise model, can debug on an enterprise scale, delivering a complete project where all you have to do is receive the finished work. The "Ultra Code" feature is where you ask Claude a question, and it dynamically plans the process using scripts, codes all the agents for you, and launches what's known as parallel processing. Multiple agents and sub-agents will coexist, exchange information, have different perspectives, and work in parallel to analyze code, perform migrations, conduct tests, perform analysis checks, validate, refute, and converge viewpoints. They are capable of stopping a task, resuming it, and running autonomously on the same job for up to 10 days. Can you imagine the amount of context management these systems handle? It's simply astronomical. This is the revolution offered to us by Claude Opus 4.8. But the problem is that all of this, the fact that they deliver a... The final result relies on context management and the ability to manage reasoning chains. And that's where I discovered the biggest flaws that Anthropic didn't want to reveal. In the official documentation, the Claude Opus 4.8 card system, where you have all the tests, there are 244 pages. I read them for you, and here's what it says: the Claude Opus 4.8 model doesn't improve its answers by increasing its reasoning or reasoning time. It's very clear. So, how is Anthropic going to justify the use of the "Ultra Code" feature? The entire architecture rests on these two factors: managing windows with 1 million tokens. If we get a score that isn't given, I can assume the result isn't good. And if the documentation tells us that the model doesn't improve its answers over very long reasoning sequences, then let me tell you, these "Ultra Code" features are going to cost you dearly. But be careful, it's not all bad; there are also positive aspects. You have to know when and how to use it. First, the benchmark confirmation for reasoning is exactly the same for mathematics. More reasoning, no better answer where previous models managed to be better. So, where is Claude Opus 4.8 better? They tell us that it makes four times fewer errors in identifying flaws, particularly in the code domain, than previous models. In other words: you don't have a diagram here. Do you know why? Because it means that Claude Sonnet 4.6 and Opus 4.7 make four times more errors than Opus 4.8. And yes, it wasn't very pretty to include a diagram where you have four times as many errors as the new model. It shows us that the previous models made a huge number of mistakes, and that shocked me. This information... Claude Opus 4.7 was capable of lying twice as much as Claude Opus 4.8. If you give Claude Sonnet 4.6 a task, it's capable of lying to you, saying it completed the work when it did n't. Did you think you could trust Claude, or artificial intelligence in general? Well, you're about to discover that until now, you've been working with models that tend to lie to you. The positive point: Claude Opus 4.8 is the most ethical model in its series. It wo n't flatter you, it will tell you what it thinks, it will very rarely hallucinate, and it won't engage in fraudulent behavior. But this also has very important consequences. The company Laboratoire developed a benchmark, Vingch 2, designed to test whether AI can manage businesses. In previous tests, Claude Opus 4.7 achieved an incredible, truly stratospheric score. The model blew away all other AIs, earning $11,000 by managing a business entirely on its own. While others stopped at $5,000 to $6,000, Claude Opus crushed everyone. But Anthropic hadn't revealed the truth: Claude Opus 4.7 turned out to be the most deceitful AI in the entire Anthropic series. The model started lying to users and customer support to achieve its objectives. But unfortunately, this information never appeared in the official marketing materials for either Claude Opus 4.8 or Claude Opus 4.7. That's why I tell you: be careful with companies. They're the same ones who deliver the products, handle the marketing, and tell you, "Take it, buy it, it works well, go ahead, green light." As a result, you suddenly discover that the performance on which they built your reputation and reliability was primarily a deceptive use of the model. And that's why Anthropic reacted so quickly. But I think that for all professionals, Anthropic should have They should communicate immediately by alerting us. Tell me what you think, but personally, I find it a lack of transparency on their part. The positive point is Claude Opus 4.8. It's a model that doesn't tend toward sycophancy, that won't flatter you, that will hardly hallucinate, and that is capable of identifying subtle differences in documents. So, it's a very fine model in terms of analysis. However, it's not necessarily better than its predecessors on a number of benchmarks. Although the model is the same price as Claude Opus 4.7 ($5 entry per million tokens and $25 exit), we don't necessarily get exceptional results on certain benchmarks for general knowledge or omniscience questions. It doesn't necessarily perform better than the previous model, neither on "hard" questions, nor on simple questions, nor in the "omniscience" section. We've gained in reliability, but not necessarily in the quality of the response. It's a very difficult balance to strike. With AI, we realize: is an AI capable of working reliably? Is it capable of answering you while taking the risk that, the more correct answers it gives, the more it tends to lie to appear correct? And that's what this part of the study reveals, showing us that Grok 4.2 and Grok 4.3 are perhaps the worst models in terms of behavior. On the other hand, Claude Opus 4.8 is extremely reliable, and Gemini 3.5 Flash, surprisingly, isn't particularly reliable in aligning the model's behavior and, along with Grok 4.3 and 4.2, is among the AIs with a high rate of sycophancy. In short, a disappointment that wasn't really present in Google's documentation with Gemini 3.5 Flash, and that's where Claude Opus 4.8 comes in, becoming much more reliable. When should you use Claude Opus 4.8? It's a model capable of comparing elements, documents, and lines of information against each other with good accuracy. It's a model with very low hallucination potential. Today, it's the best model, reducing hallucination by 95% compared to others. So, when should you use it in relation to cost? I would tend to say that when you're working in law, legal matters, analysis, data analysis, or coding, it's a model that offers four times the accuracy of previous ones. But let's get back to the "Ultra Code" feature. To understand when and how to use it, Anthropic provides advice and guidelines: the dynamic workflow. The idea is to search for bugs on a very large scale, where Claude Code will launch up to hundreds of agents in parallel. The key feature is that it will manage hundreds of working agents all by itself—that's the promise. To launch this "Dynamic Workflow" function, to orchestrate sub-agents on a large scale with dynamic workflows managed by Claude, you have two new functions. The "Deep Search" function allows you to launch a grouped workflow. You provide the line to execute, the task to perform, and the model will use advanced search functions, sub-agents, and perform a synthesis. You can also use the "Deep Search" function by asking it: "Look at this folder, this document, tell me if I should launch an Ultra Code function at the end." There are actually two ways to launch the "Ultra Code" functions. Ultra Code: You can specify the `--ultra-code effort` option, which will combine reasoning effort in XI and be automatically managed by Claude Opus 4.8, completely autonomously. This means you don't have to do anything, and in both cases, you can activate what's called workflow tracking. Imagine workflow monitoring as a console where you see the progress of your work and what stage the model is in, where it is in the process, what it has tested, which model it called, what the context window is, what tools are being used, how long the session lasted for each agent, and the number of agents. This will allow you to have a console to monitor the model's behavior. This brings us back to two more things: first, you see that we're dealing with windows of 1 million contacts. So at some point, we'll need to know the value of the window. Because if we have a low score, but today it's launching "Ultra Code" functions with 32% accuracy, there's a problem. Second, if we're managing hundreds of agents, what is the model's capacity to handle large reasoning processes? Are they better? And here, the documentation for the Claude Opus 4.8 card system shows that even if you set the reasoning to maximum, you gain nothing in terms of response performance. In the "GPQA" physical benchmark, other models continue to score better with more reasoning, but this is absolutely not the case for Claude Opus 4.8. And it's exactly the same with the "MATH" benchmark. We increase the reasoning to the maximum, we see no gain; the model remains at the same level. So, I'm going to ask you a question. What do you think? Is it worth launching an ultra-planar system? Because the ultra functions are Anthropic's new feature. They've released three or four ultra functions that will allow you to generate millions of tokens. These are workflows that will last between 40 minutes and 10 days of work, and in the end, the output must be a finished result, a completed task. Can you imagine the cost of the workload dynamics of these functions? So, objectively, I don't see any results that show the system is capable of doing it. So, the first ones who try it will tell us if they paid a lot and didn't get what they wanted. That's the doubt I have from a technical standpoint. In the documentation for Claude Opus 4.8, if you have to configure the model, the default level is High. If you're using an agentic system, it's recommended to use system XI. Remember that there's a relationship in the new Claude 4.8 model between the reasoning level and the model's ability to both provide in-depth and more verbose responses. The more reasoning you provide—the novelty of the Opus 4.8 model is that the more reasoning you give it, the more in-depth the analysis, the more verbose the responses, and the more tools it will provide. These three factors are linked in this new model. If you want the model to use tools, you need to be in High mode at a minimum. This means that if you're coding agentic systems for AI agents, you need to change the reasoning faults to at least High. For a simple chatbot, you set them to Low. But it's imperative to change these parameters; otherwise, the model won't use the tools correctly. Temperature fault parameters are not set by default. You need to set them as variables by default; otherwise, you'll get a 400 error in the interface. So, those of you coding AI agents, modify the output parameters. Effort calibration is now directly linked to the verbosity level, tool usage, and model response length. Therefore, there's an effort calibration in Opus 4.8 that considers the depth of reasoning, tools, and verbosity. There are also changes to be made to the prompt. You need to use more literal instructions. We have the option to launch agents automatically by requesting a "spawn" system and activating the "Fan-out" function. And if you want to block With this function, you specify "do not spawn agents," launching the entire task in a single response and blocking the generation of agentic systems. This is because Claude Opus 4.8 tends to generate fewer AI agents than Claude Opus 4.7, but it does so automatically. At the prompt level, we retain the structures with XML sections. It's four times more precise than previous models for code management and review. It's the most precise model for the code section. And it has a truly significant change in context management. If you need to give a model an instruction, for example, "use this tool" or "don't use this tool," it's now a model where you have to justify the reason why: "you use this tool because..." or "you don't use this tool because..." This is completely new and a real game-changer compared to everything that's been done before. So, to use Claude Opus 4.8 effectively, you need to integrate these prompt changes into your workflow. Another point: all examples must be embedded within input/output tags in XML format. They tell you this. Specifically for Claude Opus 4.8, the following sections must be included in the XML tags: "script", "instruction", "context", "input", and "documents section". One thing that surprised me is Anthropic telling us: "Define a role in the system prompt." Well, I suggest you do this, because honestly, all the studies show that it's something you should no longer do. In fact, on all the new Gemini models, they tell you: don't use roles anymore; it completely skews the AI's responses. So we're going to give this prompt to Claude Opus 4.8, which I've put in the interface (it's here), and I'm going to ask it to analyze this particular section of Anthropic, because I want to get the response from the Claude Opus 4.8 model and compare that response with what Claude Opus 4.7 might tell me. But here's what Claude Opus 4.8 tells us, while the official documentation tells us to "define a role": "You are an expert assistant specializing in Python." Okay, personally, I absolutely do not recommend doing this, given all the reading I've done in the documentation. Look at what Claude Opus 4.8 tells us: "Assigning a role (coding assistant in Python) introduces several risks depending on the degree of genericity of the chosen role. A role that is too generic ('helpful assistant', 'you are an expert'): it doesn't provide useful information. Conversely, a role that is too specific and poorly calibrated restricts responses. A third consequence: the role can create a dissonance between the role and the task." So we have potentially negative effects of two kinds. First, there's a register effect: the role introduces stylistic verbiage. The role is a system, what we call a "thematic encapsulator." And so, he tells us: there's a thematic drift effect and an impact on the understanding and interpretation of context instructions. So you see what's being said: Claude Opus 4.8 doesn't endorse this official section of Anthropic. I personally don't endorse it either. So, once again, you should always take a step back from the documentation. Personally, tell me what you think in the comments: do you assign a role to Claude? Yes? To other AIs? Yes? How? Have you read all the documentation that's been released over the last two years on this topic, which tells us exactly what Claude Opus 4.8 will respond to? It destabilizes the models, it locks the models into systems, it's not optimized. As a result, I've really given up. I find it works much better by structuring the prompts differently. There are a few There are lines you need to know how to read in the documentation: overthinking isn't necessarily the best solution. Now, that makes me smile, because as I was saying earlier, the official documentation for Claude Opus 4.8 clearly shows that there's no improvement with more model thinking, and they say so themselves; they already said so for Claude Opus 4.6. So, the workflow system is primarily about having a control console for monitoring the work of AI agents, seeing the operational phases and the progress of the model tools. But remember: if you launch the "Ultra Code," "Ultra Plan," or the new "Deep Search" function with Opus 4.8, the question remains: is the model capable of handling 1 million tokens with high accuracy? We don't have the answer. If you have it, feel free to post it in the comments below. I'm curious to see if we'll get back the 80% success rate we had on Opus 4.6 that we've since lost. And the official documentation clearly shows that we haven't seen any improvement in response times despite an increase in the reasoning token. So, remember one thing: these companies sell you a product, they sell you concepts, but if you don't get the result after 12 hours, 20 hours, or 10 days of work, you're the one who pays the bill. It's not Anthropic, it's not Claude. They're the ones who will pocket the tokens you pay. So, for me, today, caution is advised with the "Ultra Code," "Ultra Plan," and "Deep Search" features from Claude Opus 4.8; I'm not sure they live up to the promises they're making. What's interesting is the arrival of dynamic work in action, the management of the workflow function with the management interface. That 's good, that's interesting, it will replace the loop function for those who are familiar with it. But tell me in the comments: have you already tested the "Ultra Code" or "Ultra Plan" functions? What do you think of them? How are the agent systems managed? Let me know in the comments. Personally, I don't want to burn through my data plan in 30 minutes using it. So, tell me in the comments, are you ready to give it a try? I'll see you soon, until next time. Don't hesitate to like and boost this video. I'm counting on you. See you very soon.