You saw the news, but didn't quite understand it? Let me explain. A new model, ChatGPT 6.
1, has just been released, and the question now is: which is the best model on the market to use? In this video, we'll compare Opus 5. 5, Claude Fable, ChatGPT Astra, Kimi K3, GLM 5.
3, GPT-6. 1 Sol, GPT-6, and GPT-5. 6 Sol.
By the end of this video, you'll know which model to use, when and why, and everything that's changed. ChatGPT now has three models: GPT-6 Astra, GPT-6. 1 Sol, and GPT-6 Luna.
The main point is the cost. GPT-6 Astra: $10 per million tokens for entry, $50 for exit. GPT-6.
1 Sol: $2 for entry, $10 for exit. And today, GPT-6 Luna: 10 cents per million tokens for input and 50 cents per million tokens for output. And that's exactly what we're going to try to compare: what is the intelligence capability of these models and what can they do?
The cost of intelligence is decreasing. To fully understand the evolution of GPT-6. 1 Sol, we'll compare it to GPT-6 Sol, GPT-5.
6 Sol, and GPT-6 Astra. If we limit the analysis, as OpenAI did, to this interface, we're missing crucial information : which is the best model and why? What's happening is that OpenAI is developing increasingly affordable models, with a level of intelligence above the average model.
With GPT-6. 1 Sol, we went from an intelligence level of 36 to 48. We gained 12 points while reducing production costs.
Intelligence is therefore less expensive. Comparatively, today, GPT-6 Astra is certainly the most expensive model, but not necessarily the most intelligent. From a Medium and High level, GPT-6.
1 Sol becomes the most cost-effective model on the artificial intelligence market. So, between Opus 5. 5 and Fable 5.
1, which should you choose? In the middle position, you have Opus 5. 5.
Intelligence/cost ratio. If we look more closely, today, Opus 5. 5.
. . Where GPT-6.
1 is at 50, they are therefore one point behind in intelligence, which isn't really significant. However, in terms of cost positioning, we see that OpenAI has chosen a significantly cheaper model, much less token-intensive, and this has a huge impact on the amount of work you can produce. For equivalent work, GPT-6.
1 will consume up to 7 to 8 times fewer tokens than Claude Opus 5. 5. Moreover, as I showed you in the previous video (which I'll link in the description), Anthropic's current strategy is to have models that consume a large number of tokens in order to quickly move from a fixed-price plan to overage charges and sell you API connections.
And that's exactly the positioning of Claude Fable 5. 1. Technically, we see that Claude Fable 5.
1 and Astra follow the same dynamic in terms of cost relationship. The major difference is that GPT-6 Astra will be able to perform the same level of work, but with far fewer tokens than Claude Fable. So today, OpenAI is certainly the company with the best flagship product at the moment: GPT-6.
1 Sol. And I'm going to explain why. If we compare it to the arrival of Kimi K3 or even GLM 5.
3, we realize that the gap in model intelligence is now very clear. We have Chinese models that are capped around 34-35 intelligence points, while American models are at over 50. There's another very important point: we realize that initially, we thought Kimi would be cheaper.
But when we put the work into perspective, we realize that Kimi is heading in the same direction, in terms of cost, as GPT-6 Astra and Claude Fable. The reason: its token consumption and its intelligence level of 44. What does ChatGPT 6.
1 really bring to the table? We're going to talk about benchmarks to understand the model and how it changes your day-to-day work. How should one read A benchmark?
Actually, a benchmark breaks down into three phases: the task, the evaluation, and the comparison. The task is: what work do you want to accomplish? Correcting code, reading a document, using tools.
The evaluation is: is it capable of responding reliably, respecting constraints, and using the right tools to achieve the objective? And the final point: how much did it cost you, and did it achieve the objective? This is the crucial point to grasp to avoid falling into the trap of AI marketing.
It's a common occurrence. You're promised very high scores and often exorbitant costs. Here, we'll try to unravel how to approach the model and what its level of reliability is.
Deep SWE is a software benchmark that measures whether ChatGPT is capable of advancing a real software project. This will therefore interest developers, those working with no-code, and those working with automation. Is the model capable of understanding a project, modifying the code, and verifying the result?
And the goal today is to determine if it can work completely autonomously on a project from start to finish. We'll understand the results of GPT-6. 1 Sol by looking at the evolution compared to GPT-6 Sol in less than a week, and understand which is currently the best and least expensive model on the market.
First, we realize that we've gained more than 10 points in autonomy. Technically, this is a remarkable achievement. Technically, we're very close to GPT-6 Astra, and all this while having reduced costs by a factor of 5 compared to GPT-6 Astra.
Regarding the break-even point, if you look closely, you'll find, at a certain point, a break-even point in profitability. This means that all models today experience a drastic drop in profitability when you go beyond level XI . So, if you need to configure and use it, XI will be the best compromise in terms of cost/intelligence for the most complex tasks.
The rest of the time, the High level even seems to be better in a number of cases. And what emerges from GPT-6. 1 is that XI tends to overthink and therefore penalize the model.
When you have ChatGPT Astra or ChatGPT 6. 1 in your hands, the only obsession you should have is: how am I going to convert tokens into business, tokens into productivity ? That's exactly what I did.
I'm going to show you how I developed an agent that handles taxes and manages all the invoicing and accounting. What I'm about to show you is how my agent manages all the accounting for my company. I just received an invoice, a sale, and I'm no longer the one creating the invoices.
I'm no longer handling the accounting; my agent is. I'll show you how it works. An AI agent's reasoning engine needs to be able to determine if you're dealing with a professional, and whether the invoice should be addressed with a VAT rate in Europe or outside the European Union.
And that changes everything when it comes to developing reasoning engines. And that's exactly what ChatGPT 6. 1 can do: identify each invoice and the declared VAT amount.
But you want to know what's even crazier? By coding a system with computer control, it will behave exactly like an accountant. That's what it's doing right now.
It's checking that the transaction exists in Stripe's database. It's directly connected to my sales interface. It will retrieve the payments, the dates, and verify that the names, email addresses, and amounts match the receipts, because there's one thing that will cause all invoicing systems to fail: installment payments.
As soon as you have an installment payment, the model must automatically switch to It generates down payment invoices and includes the correct legal information. And that's exactly what the model can do. Right now, we're discussing this; it takes over the screen, navigates, checks the transferred amounts, retrieves the information, updates the data, and then generates the invoices.
That's what we do today. We have an accountant in the company who manages everything and who, tomorrow, will handle the URSSAF declaration, the DGFiP declaration, and the VAT. That's what we do today.
What would a good accountant do? They would systematically check for any outstanding payments from the client. And that's exactly what the model is doing now.
It opened the payment interface and then checked for any previous payment defaults. And that's exactly what any good accountant should do. When you have installment payments, to avoid late payments, and you've already had this experience, well, now ChatGPT can tell me which client might cause problems by using the payment history.
The model will therefore match the data. We're talking about real work here. So, it won't just check that the information exists; it will verify that all the details match, because what binds us as professionals is when we're going to issue accounting documents, invoices, and tax returns.
And there can't be any errors. So the model will do exactly, and even more, what a chartered accountant would do. It will verify that the data entered into Stripe's payment database corresponds to the data in the purchase interface, so that there are no errors in the name, the invoice, or the address.
I'll show you what happens inside. The model will therefore check the payment records and history in the database and then make a decision. From there, it will call a set of skills that will allow it to generate the invoice or deposit invoice.
The model has just identified that it's a deposit invoice. This type of work used to take me 30 to 40 minutes per invoice. Now, the model is generating a preview of the invoice—formatted, finished, complete, with the deposit information and perfectly compliant with regulations, including all legal requirements .
And ChatGPT will send the invoices to customers with secure links. This is what businesses need to be able to build today. In this architecture, you'll see that there's a loop with an agent that checks the invoice.
If it's compliant, it will store it in Google Drve. And that's exactly what it's doing. It's going to connect to a database, save a backup copy, and archive it.
How long did it take to write the email, create the secure link, check the email content, and send it? Eight minutes for one email. It just sent it, so we'll see in real time: it sent it.
So, what's extremely important is that from now on, I can let it run on its own. Having ChatGPT 6. 1 work independently is possible.
What you need to understand is that for all high-responsibility tasks, you have to pause the workflow to be able to intervene, check what it's done, and then restart the second part once you've checked. So, for all high-risk tasks, I would advise you to add a pause and check step, and therefore break up your workflow between the analysis and preparation phase and the sending phase. From now on, I can shut down my computer and let the AI run.
And today, we have the option of using OT systems. You no longer need of a computer. You code your AI agent on the OpenAI server, and it's the agent that will continue to work, to run.
It has access to your tools, your interfaces, and you no longer need to physically have a computer open. That's what's changing the game in the world of artificial intelligence. You have your virtual employees working for you 24/7 , and they no longer need you to be there to perform their work.
That's how we need to think about our relationship with artificial intelligence today. But let me tell you: it won't happen with prompts like the ones you see on all the platforms. I've been explaining this to you for three years.
If you haven't understood it, the job market will make it clear. The development market will make it clear. The entrepreneurial market will tell you: you don't have the skills and experience required to use these tools, to produce real work, to free up your time, to increase your margins, to allow you to do 30 to 40% more work with a subscription.
That's what I do. So, if I have one piece of advice to give you, it's to seriously dedicate yourself to engineering reasoning systems to code system instructions that you can let the AI work with for 36 hours. Managing all the accounting, shipping, analysis, data, and finance aspects: today, there are practically no limits.
The only limit is you. You become the weak point in the architecture. If there's one benchmark you need to pay extremely close attention to in the professional world, it's this one: the benchmark that tests PDF comprehension.
Let me explain. When you look at the GPT-6. 1 Sol curve, you see a constant on the left side.
And I'll explain what that means. This trend is present in all models. Yes, but not with the same confidence level.
First, GPT-6 Sol had a confidence level of 24. GPT-6. 1 Sol has a confidence level of 30.
When you give a PDF to an artificial intelligence, you only have a 30% chance that the model will be able to understand the details within the PDF. How many times have you seen tutorials explaining that you can send entire PDFs into ChatGPT? Well, know that with the best models on the market today, we can't achieve more than 30% accuracy on relationships within a text.
You have to understand that a text is made up of thousands of relationships and concepts. That's exactly what we have here. We have a knowledge base, and the model isn't capable of relating one element to all the other blocks in the text.
That's why, when creating a knowledge base—for example, a second brain—you have to force the relationships within the document structure. It's the only way to solve this problem. In a previous video on GPT-6, I showed that even with the best models on the market today, if you tell them, "You're in the construction industry, you have all the skills to create quotes in the electrical field, give me a quote for a studio apartment," the model can't generate a quote.
Then, some of you, one or two people, replied, "But you could just give it the NF C 15-100 standard, or find an API and feed everything into it. " That won't work . That's exactly what's being explained here.
We have a 30% cap, the best model, Astra, at 40%, and Opus 5. 5, on the other hand, isn't particularly good with PDFs: it caps around 32. And the main difference is that the cost is double compared to GPT-6.
1. So, if you have to use PDFs, it's better to use GPT-6. 1 Sol.
There 's another factor as well: the level of reasoning increases the relationship and text comprehension. So, if we increase the model's reasoning level, we also increase its ability to understand the document. Each level increases the model's capacity by approximately 2%.
And we see that Opus 5. 5, on the other hand, is clearly counterproductive. Therefore, these are not models that should be used with raw PDFs by default.
Another point: the cost becomes enormous with GPT-6 Astra. So always prioritize GPT-6. 1 Sol: it offers the best cost/performance ratio.
It achieves the same performance as Astra on PDFs, but at a quarter of the price. If you want to understand how to create knowledge bases that can be used by artificial intelligence, I'll show you exactly how. In The Best of AI, you have 90 days to go from zero to professional and automate your work.
Today, knowing how to create AI agents capable of assisting you with social security contributions, accounting, communication, and analysis is the only way to improve your margins and profitability, free up time, and allow you to focus on the very few people who currently have the skills to code reliable AI agent systems for their businesses. Don't make the same mistake as those who used AI and end up among the 70,000 business failures predicted for 2026. Can GPT-6.
1 replace a junior employee? Well, GPT-6. 1 outperforms GPT-6 at using a computer, with a score of 73%.
It costs half as much as GPT-6 in terms of tokens and six times less than GPT-6 Astra. Technically, this means that GPT-6. 1 is capable of producing real work.
Have you seen the situation with invoice management? Well, that's exactly it. It 's capable of producing work autonomously, much cheaper than GPT-6, with virtually the same level of reliability, at the lowest price currently on the market.
So, there's only one goal you should have today: starting today, you begin coding your agents with GPT-6. 1, you put them into production, and by next week, you'll free up time on your workload. By reliability, we mean its ability to detect false information.
So, technically, GPT-6. 1 is extremely reliable compared to other models. However, this needs to be put into perspective.
As we saw in the PDFs, the reliability level is 32 to 34%. So, it's always important to consider the context. The quality of the data you provide is the entry point for the model's detection and understanding.
An interesting point is that they've improved the model's ability to understand complex prompts . This means it will better respect boundaries. And the boundaries we're trying to test with an AI agent are: Is it capable of lying due to a failure?
Does it bypass security codes? Does it respect restrictions? And does it stay within the limits of authorized tools?
When granting access to the payment system, Stripe, or invoicing, these four points are crucial for choosing the right model to support your daily work. If you have to choose a model for everyday use, GPT-6. 1 Sol is twice as secure as GPT-6 Sol.
GPT-6 Astra remains the best model. But above all, for those using GPT-6 Luna, don't use it for high-responsibility tasks. We have a failure rate close to 30%.
GPT-6. 1 respects restrictions three times more than GPT-6 Sol. Technically, it won't try to circumvent the restrictions you impose, and that's extremely secure.
Even if you give it a computer to control, it's capable of doing so while respecting the guidelines and without exceeding the measures you've set. Therefore, it's crucial that, in your prompts, you can clearly define the limits, parameters, and tool calls within the model's operation. If what I'm telling you seems complex right now, and you're still at the stage where you're talking to ChatGPT and saying, "Find me customers online, give me a report," then now is truly the time— perhaps your last chance—to get into artificial intelligence.
The technical level is extremely high, and contrary to what some might try to make you believe, it's anything but easy. That's precisely why I've developed my content, videos, and training programs: to give you access to best technical practices so you can understand which model to use, when to use it, and why. And if you have to choose a model today, switch to GPT-6.
1 right now. You have the most balanced, most affordable model on the market, even compared to Chinese models, with the most consistent scores for 99% of users. And those currently involved in scientific research will have access to GPT-6.
1, which is optimized for data analysis, simulation, and mathematical analysis. They've developed a system for thought processes and scientific reasoning, allowing the model to double its reasoning score and approach that of GPT-6 Astra, while reducing labor costs tenfold compared to GPT-6 Astra. We certainly have the best model of the year here.
Feel free to share your thoughts in the comments, whether you've tested it, code with it, or have already created AI agents. Share your best tips in the comments below. See you soon.
Until next time. Subscribe if you haven't already.
