Claude 4.7 : les 3 nouvelles fonctionnalités qui changent tout

Claude Opus 4.7 has just been released. Here's what's new. In this video, we'll talk about the brand-new features with the Ultra system, the new effort function, the changes to Claude's configuration, its architecture, and the introduction of new commands. How do you use them? When do you use them? I'll tell you everything in this video. So, don't expect to use Claude Opus 4 at all. We're dealing with a change in behavior here. Anthropic is clearly warning us: we're no longer using the same operating system. And for good reason: it's a model that makes decisions, but if you don't guide its decisions, well, you don't really know what it's going to choose. The most aligned model today is Anthropic's Mythos Preview, which is certainly the next benchmark and the next architecture to use. A misalignment can also manifest as this type of problem: I give Claude a very long prompt, and it actually has trouble following the instructions. And this is something the Anthropic team has really improved. The first behavioral change in this Opus 4.7 system is a literal adherence to instructions. This means that what you write, the model will do, but what you don't write, the model won't do. Remember those presentations and prompts that said, "Give me a pitch deck presentation to sell my product," and you had Claude starting a PowerPoint presentation? Well, what it's going to do now is wait for you to give the instructions. And I'll explain why. The entire artificial intelligence industry is trying to solve two problems. Is a model capable of working and therefore thinking? Apple has amply demonstrated this: these are models that are not capable of thinking. So what happens is that at some point, it loses focus on actions, and that's what we want to avoid. So what we're developing are systems in which we 'll code instructions. We'll tell it what to do, and that's exactly what the model allows us to do today. We'll develop the instructions that will allow the model to navigate the gaps, and for this, they have two strategies. The first strategy is to have created what's called a model with an Advisor function. You type "Advisor" in your interface, press Enter, and in the terminal, you realize that you can actually link your Claude Opus, Sonnet, or other model with another model, whether it's Sonnet or Opus. That is to say, every time you send a request, if your model is stuck, it will ask another model (in this specific case, Opus 4.7) how to solve the problem. And Opus will send a very precise sequence, and it's extremely important to understand this: 400 to 800 tokens, so it's a very short sequence to tell it how to solve the problem and how to act. The Advisor Tool function was released 48 hours before the new model, and as I predicted, it's the new agentic architecture. The AI ​​no longer works alone; it always works in tandem, meaning you have at least two agents. This system allows for greater stability in the models' reasoning, and it actually allows one model to restart the other when there are roadblocks. What doesn't change are the instructions. If we can't code the instructions, the models won't know how to proceed. They'll communicate with each other, but they won't know how to do it. That's why you need to get involved in writing the internal prompts. I'll give you an example later in the video. To briefly remind you, agentic systems are essentially systems where you have what's called a "head agent," meaning a conductor. In this specific case, it's Opus 4.7 that will have to coordinate agents. These agents are what we call "sub-agents." These sub-agents have their own functions, contexts, and isolated tools. That is to say, they have their own tools, their own operating systems, and they will therefore use their configuration and send their work to the head agent. If you prefer, they are employees that you are configuring. But to do this, let's be very honest with ourselves: it's not going to happen automatically. Forget all the promises that tell you, "Okay, you ask Claude, and Claude will do all your agents by himself." Okay, that's good from a marketing perspective, but it's not true in the field of entrepreneurship. What the model knows how to do is follow instructions, and that's what we ask it to do. I'll give you an example with a system that has what we call an orchestrator system. I have a model that is an orchestrator that will distribute work to sub-agents. This means that the system I coded is called an "orchestrator." It acts as a pilot and has employees called "agents"—the famous AI agents. What Claude needs to understand is what we call the internal architecture of the model. And that's exactly what we need to convey to the model's operating mode. It's called an ASCII decision tree. And that's exactly what you need to do to make Claude 4.7 work correctly. You have to allow it to perform the crucial step: the agentic loop. Your instructions must enable it to understand "what I need to retrieve," "who I need to work with," and "how I need to choose my actions." Therefore, your instructions aim to quickly make the model understand what it needs to do. In other words, a decision-making path. This is exactly how we'll test if the model's behavior is aligned, meaning if it follows the instructions and maintains long-term consistency. So, your main agent, the head agent, the orchestrator, will query other agents. These agents will be, for example, the work system, which itself has its coded instructions: the code, the model, the effort level, its role, what it must do in its operating mode , which variables it must work with, and how it must format them. Then, it will retrieve this information, send it back to the head agent, which will compile all the data and deliver the result. Now, you understand one thing: agentic systems aren't a simple "do my job" prompt. If you want to model your office work, this is exactly what you need to do. First, because AI doesn't know how you work in your office or in your profession. So, you have to code the instructions precisely for it; otherwise, you have a model that keeps thinking. So, agentic models will continue to produce tokens, but they won't produce a solution. The blue and purple bars are bars of proposed solutions. So, you see, the models ultimately struggle to perform valid work, and we have to code these instructions. If we don't, we'll always have unusable content. That's the first point we're trying to solve, and Claude and Anthropic have solved it perfectly. That's why the biggest American companies today are able to automate their jobs and are laying off staff left and right. Today, I'm going to be very honest with you. I think I've created one of the best training programs on the artificial intelligence market, and I've decided to raise the prices for a very simple reason. Today, I've attended dozens of training sessions where the instructors are guys who create perfect prompts. I spent another €5,500 on training last month and I've come to a conclusion. Technically, the skills I'm giving you are so far superior, so far superior, that it's worth it. Ten times the price I'm offering you for this training today, but significantly so. Those who want to learn can still do so at the advertised price until next Saturday. After that, I'm raising the training prices for a simple reason: today, I'm not just giving you prompts, I'm giving you the strategy to automate and professionally scale your business using artificial intelligence. I'm not just showing you ChatGPT or Claude; I'm showing you ChatGPT, Claude, Gemini, Genspark, the systems for creating presentations, chatbots, voice chatbots— in short, you have a complete system with prompts and methods used in the professional field. All the information is in the description. Anthropic reveals that Opus 4.7 is a "Light" version of Mythos preview. Why? Because Mythos is primarily designed for cybersecurity. So, Opus 4.7 is a cyber-light version with restrictions to prevent it from being used for massive attacks. But they've improved quite a few benchmarks, especially the coding and reasoning aspects, and some tools have been significantly enhanced. The gradual reasoning aspect, based on academic knowledge, is approaching the perfection of human knowledge: 94.2, roughly speaking, meaning it knows almost everything and is capable of reasoning. But there's one point, a really crucial point, and that's visual reasoning. So, here's what it changes. To be honest, if I sent a graph like this to Opus 4.6 or Sonnet 4.6, it didn't handle it very well. That's the real problem, what we call visual reasoning. Now, that's fixed; it's among the best models. They are therefore capable of understanding multimodal graphical technical elements and linking them to the text. What Gemini 3.1 already does very well has been fixed in Opus 4.7. In general, Claude Opus 4.7 shows significant improvement in its coding performance, particularly in benchmarks. It's noticeable that they've invested in computer usage. As you know, Claude can now take control of your computer, and this update is a step in that direction. We'll increasingly delegate tasks where Claude can take control of your bash command and execute complete workflows. One area where performance has declined is agentic search. While not dramatic, there has been a surprising drop. I don't find it particularly worrying, but it does mean that we might need to pay a little more attention to the automated search of internet content. For those of you using the classic interface, the online chat interface, you simply have the Opus section with a feature called "adaptive thinking." This system automatically manages the model's reasoning capabilities. Later in the video, I'll explain what's changed in the model's reasoning, because there have been quite a few modifications. Regarding skills, this model has GDPVAL skills. If GDPVAL isn't familiar to many of you, I'll give you a brief overview. Essentially, we've modeled behaviors. It's similar to the concept of skills. In fact, Claude has released a brand new Claude Cowork interface with a new menu. We'll talk about this in another video, but to give you an idea, GDPVAL functions are functions where we've taught "how you're going to create a PowerPoint file," "how you're going to create a Word document," and we've given it skills in different areas: vision, information reasoning, long-term vision for complex tasks, science and biology, and long-term consistency—that is, is it capable of making decisions? Because What you really need to understand—the shift that has occurred, which has completely changed the world of work and will mean that we're practically all going to be unemployed—is that Claude is an agentic system in itself. So, you shouldn't really use it like a chatbot anymore. In any case, if you're still using it as a chatbot, you're losing 98% of Claude's power. The agentic system breaks down your prompt into instructions. These instructions must be followed. This is what we call the behavioral alignment of a model. The instructions you give will define what we call an agentic loop. A loop is a context retrieval, an action taken, an iteration, and this loop can last for a certain amount of time. And it's precisely this context with what we call the "long-term coherence benchmark 2." It's a system that will allow Claude to run and make decisions. For example, we'll give it customer support and let it run and see how long it can manage a business on its own. And in this specific case, it manages to generate roughly $10,000, all by itself. The Opus 4.6 version generated around $8,000. In conclusion, we've seen progress. It's more autonomous, it can work longer. All of this is thanks to what we call instruction alignment. And this is extremely important because the Opus 4.7 model has a misalignment score of 2.75. Opus 4.6 is slightly lower, around 2.48. So there's a slight improvement. Misalignment is when a model decides to do something other than follow your instructions. And this is very important because, remember, we've entered an agentic system. If you have an agent that's acting erratically, well, by the time you can get it back, it'll be too late. There's a technical point we need to address: Claude's reasoning. Claude now has a new reasoning system called "Extra High," and Anthropic has given us some very specific instructions for using it. We also have a new command called "Ultra." I'll tell you about it because it's simply amazing, but I'll also show you what it costs to use this type of function. Anthropic tells us that for Claude 4.7, the recommended operating mode is "High" almost all the time. I 'll show you the graphical results later. The "Extra High" system is recommended for the code and for tasks lasting more than 30 minutes. So, as soon as we're dealing with analysis, data retrieval, or code development, we're talking about budgets that will run into millions of tokens. And that's the configuration recommended by Anthropic on Claude 4.7. Now, I know many of you are going to say, "Yeah, that's going to be super expensive." But I completely agree with you. Anthropic has taken a gamble: talking to businesses. Today, they're not trying to... in my opinion, to quibble over price. You want quality, and that's pretty much my point of view too. You want quality, and that costs money. So you have to invest in quality at some point. If you want something halfway decent, well, you can use Chinese models, or models from here and there. They'll never have that level of technical expertise. You'll need a Claude Opus Pro or Claude Max account. Luckily, there are still a few left on the platform. The platform is at this address. You'll find the accounts and the address in the "Receive promos, AI basics by email" section. I'll give you all my tips: where I get the best rates to get my AI subscriptions at the best price. All the info is in the description. And now Claude Opus 4.7 is available. At least to optimize the process. Based on the results of the studies, we can see a significant 10-point gain between Sonnet/Opus 4.6 and Opus 4.7 in the High configuration. But I haven't told you everything. The Extra High version offers performance equivalent to the maximum of Opus 4.6 at a lower cost. So, in the end, Extra High in reasoning is equivalent to 110,000 tokens. This means you absolutely need a Max plan with a one million token window. The plan I advise against is putting it in Max mode. Why? Because while you gain 3-4%, you go from 100,000 tokens to 210,000 reasoning tokens. That's considerable, and for me, 3% isn't worth it. So, the configuration is High/Extra High for 99% of the work with Opus 4.7. If we're in Medium mode, we're barely at 55% reliability, which is too low. So in High mode, we start getting more reliable answers. And I'm going to give you my perspective on why this problem exists. From experience, I'd say there's a relationship between the number of layers in the neural network and the number of attention blocks. Think of it like the width of a river and the height of its walls. If we give the model a lot of information, if the model can't handle the enormous amount of data, it overflows. That's the number of attention blocks. If the model has more attention blocks, then it can contain and manage a larger amount of information. Then there's another point: how it handles a large volume of data. And there, as I said, that's the depth, or rather the breadth, whichever you prefer, of the neural network. You know, we always come back to what I told you. What we're trying to solve today is how to send a lot of tokens without the model collapsing. These are always the two points we try to solve in architectures. So what we're implementing are neural systems with greater breadth and depth: attentional blocks, neural layers. As a result, we can handle a very large number of tokens while maintaining the relationships between concepts. If you haven't seen them, I'll put the videos where I explain the problem in the description. Is an AI capable of correlating and understanding what you're asking in a text of a million tokens or 500,000 tokens? Well, not really. You realize that the models lose focus, and that data quality has an impact. When you start increasing the amount of information requested within a text, well, we see even more model breakdowns. This is where the problem lies: the model doesn't understand what to search for and how to use the information. The method—well, they increased the number of neural layers, they increased the number of attentional blocks—but that doesn't mean you should provide poorly structured data. How do you activate all these parameters in the Claude interface? In the Cowork interface, it's not possible to modify the model's reasoning level. It automatically operates on an "adaptive thinking" system. Claude manages this on the servers. However, everything changes when you 're using the CLI-type interface. Don't forget to subscribe and turn on notifications because I'll probably be releasing videos to teach you how to use Claude CLI. It's currently the most perfect system for businesses. The technical level required to work with it is simply incredible. You control all the AI ​​parameters. So if you haven't already, subscribe and turn on notifications. As soon as it's released, you'll have the new video to help you use and configure Claude CLI for your business. And don't forget to like, share, and give me some support in the comments. It's super important for SEO. So, first thing, if you choose a template from there, you'll type the "/model" function. Then you'll type... The name of your model, and in this case, for example, I can select a Sonnet, and we'll automatically switch to Sonnet 4.6. If you want to switch back to an Opus model, you'll type the Opus version, and it will automatically switch back to Opus 4.7. If you want to see all the available models, type "/config", go to the config section, then to the models section, press the space bar, and there you'll see the models currently available by default. Now that you've chosen your model, there's another factor: the recommended effort level we discussed earlier. Are we going to put this little Opus through its paces or not? This is where you'll type the command "/effort". And there you'll choose. And you'll see that you now have a new parameter: automatic, max, and extra high. This is where you can switch the model's effort mode and choose. It's truly the Rolls-Royce of software. You configure everything just like an engineer would. So, all you guys who do development, data analysis, and even office work, don't be fooled. Office work is simply the ultimate tool, Claude. Claude Code in its CLI version is the workflow-building machine for businesses. Regarding the model writing section, it has a more direct tone, a different writing style, and you'll probably need to create "custom instructions." So, where do you go to do that? I'll show you where to configure it because the model length and the response method are very different from Opus 4.6. In the config section, if you go to the config section in Claude CLI, you'll find what are called "output styles." And by pressing the space bar, you have different default modes here: "explanatory", "learning", and I've added one more, which is "code_style". So, to create this type of system, you go to a section called "output style". Here, right in the directory, you create a file. I'll show you how the files are created. You define a name for your system, you give it a description and style, and then you specify how the model should write. If you like, it's an improved version of "custom instructions" by specifying the tone and the target audience. In short, it determines the sampling of your model's output tokens . And then, you select your model, and the model will write in the way you defined by setting its probabilities using the "output style" instructions from Claude Opus 4.7. We're going to talk about the new function called "Ultra Review", which will replace the "review" function. It's a complementary system. So, I 'm telling you about it, but be aware, it costs almost 200,000 tokens. It's an automated system that will send AI agents to review your code with three independent agents, and for all Pro and Max account holders, you get three free runs during this period. So, if you want to test it, now's the time. It's a system that will perform a stability test, check for system vulnerabilities, fix bugs—in short, it's a complete review. How do you activate it? I 'll show you how. When you're working in the interface, in your workspace, you activate the Ultra Review function, which is right here below, specifically "/ultra_review". And they tell you, the cost of using this function varies between $5 and $20. So expect it to take a good 40 minutes— a full 40 minutes of work. You select the section, define the working directory ( i.e., the directory area), drag the link, and let the AI ​​agent do its work. This will result in an analysis file and bug fixes for your software. We're also seeing the introduction of a function called the proactive function. This proactive function is very similar to what we call "loop" functions. These are automated systems where you type "/proactive" in the menu, and what happens is that prompts are executed according to a schedule. For those using loops, you specify an interval, for example, 5 minutes, and then you provide your prompt. Every 5 minutes, you display some information, and then you type your prompt again, and the system will handle this in a loop. It will create what's called a "cron task," an automated task. Generally, this is limited to 3 days. So if you want it to run for longer , you have to specify it in the system. But this allows you to launch automatic loops with the function, and here's an example: "/loop 5 minutes. Check if the deployment of my app on GitHub has been completed." So it's a truly automated task that you can set up, and they've introduced a new system called "Proactive." My final piece of advice regarding Claude 4.7 is that the prompts that used to cause problems, those with too many sequences, are the ones you should now use. Not long sentences, but prompts that detail exactly what you want. Also, don't use Claude with version 4.7 with reasoning below "High." It's not recommended; the results aren't very good. And yes, there's also the point we discussed: image resolution. It can now handle high-definition images, hence its improved understanding of graphs. They've also released a new function, the "/undo" function. They regularly change its name. It's "/rewind." It lets you go back one point in a conversation. Let's say that at some point, you encountered a problem with your configuration, in the interaction with the model, and you need to revert to a previous state. You type the "/rewind" function and restore the code or a previous conversation. You press Enter, and you'll have access to all the different conversations. This is the current conversation, and if you scroll back, you'll find other conversations, allowing you to go back. This is useful when you see that the model has strayed into a discussion, so you avoid cluttering your current conversation. As you've seen, I have a real-time streaming system to see how many tokens I'm sending and what's being cached on the Anthropic interface. If you'd like us to make a tutorial on how to install this type of interface, let me know in the comments. Don't forget to subscribe, like, share, and comment. And for those who want to find me, I'm now on LinkedIn. So, see you soon. Until next time.