One of the three AI models in this comparison finishes a benchmark task for about one fifth of the cost of the top model on the chart. Its weights are due to be downloadable on October 27. And it is named after a cat meme.
That model is Mistral Large 4, better known as Le Chonk. Everybody is asking the same thing: is it actually good, or is it just funny?
The honest answer has two halves. Le Chonk is a real achievement. It is also clearly behind the two American flagships on raw intelligence. Both things are true at once, and which half matters to you depends on who you are.
So I lined it up against Claude Opus 5.5 and GPT-6 Astra using independent numbers, not marketing slides. No hype, no panic, just a clear look at who wins where.
One note before we start. AI benchmark scores move around. Test versions, effort settings and providers all change the results a little. Where sources disagree, I will tell you.
What You Will Find in This Post?
- The quick answer for people in a hurry
- Meet the three models
- The independent scoreboard
- Where Le Chonk wins
- Where Le Chonk loses
- Claude Opus 5.5 vs GPT-6 Astra: the effort dial changes everything
- The real cost per task
- Who should actually care about Le Chonk
- What to watch on October 27
- Frequently asked questions
- Final verdict
1. The Quick Answer (Read This If You Are in a Hurry)
If you only have 30 seconds, here it is.
- Smartest of the three: Claude Opus 5.5.
- Cheapest of the three: Mistral Large 4.
- Most control over your data: Mistral Large 4, once the weights arrive.
- Middle of the pack on score, most expensive per token: GPT-6 Astra.
The scores below come from the Artificial Analysis Intelligence Index, and the Mistral and Claude token prices match OpenRouter’s side-by-side comparison.
| Mistral Large 4 (Preview) | Claude Opus 5.5 | GPT-6 Astra | |
|---|---|---|---|
| Maker | Mistral AI (France) | Anthropic | OpenAI |
| Weights | Open, due October 27, license not yet published | Closed | Closed |
| Intelligence Index score | 38.4 | 57.6 (max effort) | 52.7 (max effort) |
| API price per 1M tokens (input / output) | $0.68 / $2.09 | $4 / $20 | $10 / $50 |
| Cost to finish one index task | $1.13 | $5.98 (max effort) | $3.26 (max effort) |
Prices differ by provider, so check the current rate before you build anything. That one table is the whole story: much cheaper, much less capable, and wide open on control.
2. Meet the Three Models
Before the numbers, a quick introduction to each player.
Mistral Large 4 (Le Chonk): The Open Challenger
Mistral Large 4 is the biggest open weights release ever from the French company Mistral AI. It has about one trillion parameters, it is multimodal, it reasons, and it uses a mixture of experts design with about 49 billion parameters active at a time.
That “active parameters” detail matters. The model is huge on paper, but it only wakes up a small slice of itself for each token. That is how it keeps serving costs low.
Mistral opened a public preview on October 6, so you can try it through the API right now. The weights come later.
There is one catch you should know about early. The license terms have not been published yet, and nothing guarantees they will match the Apache 2.0 license that the earlier Large 3 used. Until that text is out, “open weights” does not automatically mean “free to use for anything”.
Want the full story on this model first? I broke down the one trillion parameters, the 49 billion active ones and the October 27 date in Mistral’s “Le Chonk” Has 1 Trillion Parameters but Uses Only 49 Billion: 10 Things to Know Before the Open Weights Land.
Claude Opus 5.5: The Current Leader
Claude Opus 5.5 is the top model from Anthropic in this comparison. It launched on September 22, and Anthropic priced it at $4 and $20 per million tokens, which is 20 percent below Opus 5.
Here is the twist, and it comes back later in this post. Even with the lower token price, Opus 5.5 can cost more per task at the top setting. At maximum effort, it burns roughly 119,000 output tokens per task, which pushes the cost to $5.98 per task against $3.26 for Astra.
In plain words, a lower price per token does not always mean a lower bill.
GPT-6 Astra: The Expensive Reasoner
GPT-6 Astra is the high-end model from OpenAI in this group. At maximum effort it scores 52.67 against 57.62 for Claude Opus 5.5 on the same index version.
You may see other Astra numbers online, such as 54.7 or even 61.2. Those come from different index versions and earlier runs. When you compare models, always make sure the numbers come from the same version of the test.
Astra is also the priciest per token. Its rate card runs $10 and $50 per million tokens, against $0.68 and $2.09 for Mistral Large 4.
Also comparing the other big names? I put Grok and GPT-6 Sol on the same table in Claude Opus 5.5 vs GPT-6 Sol vs Grok 4.7.
3. The Independent Scoreboard
Time for the numbers. Here is how the models line up on the same test.
| Model | Intelligence Index score |
|---|---|
| Claude Opus 5.5 (max effort) | 57.6 |
| GPT-6 Astra (max effort) | 52.7 |
| Gemini 4 Argon | 52.6 |
| Mistral Large 4 Preview | 38 |
I added Gemini to the table so you can see the wider picture, not just our three.
The first independent test of Le Chonk only half backs up what Mistral said. Mistral Large 4 scored 38 on the index. That makes it the strongest open model outside China, but still behind seven Chinese models.
So Mistral’s headline claim holds up, with a very large asterisk. The model does beat GLM-5.2 and DeepSeek V4 Pro, the two models Mistral pointed to at launch. But it sits more than 19 points behind the best closed models, which is about two thirds of the Opus 5.5 score.
That sounds harsh, but think about what you are comparing. A model you will be able to download and run yourself is landing within shouting distance of systems that cost far more to use.
Claude vs GPT: Closer Than You Might Think
The two American flagships are much closer to each other. On the separate Epoch Capabilities Index, Claude Opus 5.5 took the top spot by less than one point over GPT-6 Astra.
Different test, different scale, same story. Opus 5.5 leads, and Astra is right behind it.
Curious about the hype around OpenAI’s model? I separated the real claims from the marketing in GPT-6 Astra and the AGI Claim: What Is Real and What Is Marketing.
One warning about charts like these. A single score hides a lot. The same model can look brilliant on one test and ordinary on another. That is why the next two sections matter more than the table above.
4. Where Le Chonk Wins?
Now the fun part. Here are the places where this big orange cat genuinely earns its crown.
Win 1: The Price Is Hard to Beat
Per token, Le Chonk plays in a different league. Compare the rates from the table in Part 1. On input it is about six times cheaper than Opus 5.5, and on output it is nearly ten times cheaper. Against Astra the gap is even bigger.
Per task, the saving is smaller but still large: $1.13 against $5.98 for Opus 5.5 at maximum effort. If your work does not need the very top score, that adds up fast.
Win 2: You Can Run It Yourself (Soon)
This is the real headline for many companies. Open weights models are billed only for the compute you use. There is no license markup and no vendor that can change the price overnight.
For a bank, a hospital or a government office, there is a second benefit. Your data stays inside your own walls. Nothing has to be sent to someone else’s cloud.
Keep the license question from Part 1 in mind. Read the final terms on Hugging Face before you plan a product around it.
Win 3: A Serious Cybersecurity Pitch
Mistral is proudest of its security results. Before the weights ship, the company is red-teaming Le Chonk with cybersecurity leaders, vetted partners and state authorities, who get access with reduced moderation and expanded cyber capabilities.
The logic is easy to follow. Co-founder Guillaume Lample argued that the cyber defense capabilities will help enterprises and governments defend themselves against threat actors who are jailbreaking closed models. And if a company loses access to a closed tool in the middle of an incident, that loss is a risk by itself.
A fair warning, though. These are the company’s own claims, and independent testing will matter more once the weights are public. AI that acts on its own, whether for defense or anything else, needs careful limits, as I explained in ChatGPT and Claude Can Now Click Buttons for You: 10 Things You Must Never Hand Over to an AI Agent.
Win 4: Seeing Pictures Well
Here is a surprising one. On a visual grounding test called Dense 200, Le Chonk reportedly performs slightly better than GPT-6 Astra. Visual grounding means finding specific objects inside an image.
Treat this as a promising sign, not a final verdict. It comes from the company’s own testing, and one test does not make a model better overall.
Win 5: Europe Finally Has a Real Contender
Most of the strongest open models come from China, and the rest of the open field is thin. Mistral calls Le Chonk the strongest open weight model developed outside China by a substantial margin. The independent score above supports that claim.
For European companies that care about data rules and independence, that matters. It gives them a serious option that does not depend on an American or Chinese lab.
So Le Chonk has real strengths: price, control, security focus, vision and a European home. But a champion with no weak spots does not exist, and Le Chonk has some real ones. In Part 3, we look at where it loses and why the “cheap” label needs a closer look.
5. Where Le Chonk Loses?
Every model has weak spots, and an honest comparison has to name them. Here are the places where Le Chonk falls short.
Loss 1: The Raw Intelligence Gap Is Real
We covered this in the scoreboard, but it deserves a plain sentence. Le Chonk scores 38 where Opus 5.5 scores 57.6. It beats every open model from the US and Europe, yet it still sits behind a long list of Chinese open models. If you want to understand why that race matters so much to governments, I covered the politics in Trump’s AI China Race: What the Government Is Really Trying to Win.
Reading tip: the gap is smaller than it looks for simple jobs like summaries and support replies. It is largest on hard, multi-step work.
Loss 2: Coding Agents and Terminal Work
This is where the gap shows up most clearly. On the Vals Terminal-Bench 4.0 test, Le Chonk scored 22.7 percent and ranked 21st out of 45, while Claude Opus 5.5 scored 65.2 percent. Other test runs put Le Chonk in the high 20s, but the picture is the same.
The story is not all bad. On a software engineering benchmark that resembles SWE-bench, Le Chonk beat most open models and trailed only Kimi K3. So it can fix real coding issues. It struggles when it has to run long chains of commands on its own.
Mistral admits the same thing. The company says the model still lags behind the frontier in coding.
Loss 3: It Talks Too Much
Here is a detail most headlines skip. To finish the Intelligence Index, Le Chonk produced about 200 million output tokens against a median of 81 million for comparable models. Artificial Analysis calls it very verbose.
Why does this matter? Reasoning models bill their thinking tokens as output. A model that thinks and writes a lot can turn a cheap price per token into a not-so-cheap price per task. We will do the math in section 7.
Loss 4: The Preview Price Will Not Last
The $0.68 and $2.09 rates I showed you earlier are a launch deal. Mistral’s pricing lists Large 4 at half its list price during the preview, and the full rates are $1.36 and $4.18 per million tokens.
That is still cheap next to Opus 5.5 and Astra. But if you plan a budget, use the full price, not the discount.
Loss 5: Open Weights Come With Homework
“Open” sounds free and easy, but it comes with chores.
- The weights are not out yet, and the license is not published.
- A one trillion parameter model needs serious hardware to host yourself, even with only 49 billion parameters active.
- Once weights are public, anyone can download them, and the maker cannot take them back. Safety rules built into the model can also be stripped out by other people.
That last point is why the debate over open models is so heated. AI systems that act on their own can already push past their limits in testing, as I showed in AI Agents Are Escaping Their Sandboxes: What Happened and Why It Matters.
Fair balance: closed models can fail too. Open weights simply move the responsibility from the lab to whoever runs the model.
6. Claude Opus 5.5 vs GPT-6 Astra: The Effort Dial Changes Everything
Now let me settle the fight between the two American flagships, because the headline scores hide the best part of the story.
Both models let you choose how hard they think. That setting changes the score and the cost a lot. Here are the independent numbers, from a breakdown of Artificial Analysis data by effort level.
| Effort setting | Claude Opus 5.5 (score / cost per task) | GPT-6 Astra (score / cost per task) |
|---|---|---|
| Low | 42 / $0.55 | 46 / $0.82 |
| Medium | 51 / $1.34 | 50 / $1.54 |
| High | 54 / $1.82 | 51 / $1.73 |
| Extra high | 56 / $3.46 | 52 / $2.31 |
| Max | 58 / $5.98 | 53 / $3.26 |
Read that table slowly, because it explains a lot.
- At low effort, Astra is smarter. It scores 46 against 42 for Opus 5.5.
- At medium effort, they are level. Opus is slightly higher and slightly cheaper.
- At high effort, Opus pulls ahead. At $1.82 it scores 54, which is already above the best Astra score of 53 at $3.26.
- At max effort, Opus costs a lot more. You pay $5.98 for the last few points.
So which one is better? It depends on the dial. For most work, Opus 5.5 at high effort gives you more intelligence per dollar than Astra at any setting. For quick, light jobs, Astra at low effort holds up well.
Want the other side of the OpenAI story? I looked at how the labs are fighting each other in Anthropic vs OpenAI: Inside the AI Civil War.
7. The Real Cost per Task (My Own Math)
Now let me combine the price, the score and the talkativeness. I divided each model’s Intelligence Index score by its cost per index task, using the figures above. This is my own calculation, so treat it as a rough guide and not an official ranking.
| Model and setting | Score | Cost per task | Points per dollar |
|---|---|---|---|
| Mistral Large 4 | 38 | $1.13 | about 34 |
| Claude Opus 5.5 (medium) | 51 | $1.34 | about 38 |
| Claude Opus 5.5 (high) | 54 | $1.82 | about 30 |
| GPT-6 Astra (medium) | 50 | $1.54 | about 32 |
| GPT-6 Astra (max) | 53 | $3.26 | about 16 |
| Claude Opus 5.5 (max) | 58 | $5.98 | about 10 |
Here is the surprise. Le Chonk is cheap per token, but it is not the best value per point once you count how much it talks. Claude Opus 5.5 at medium effort scores 13 points higher for about 21 cents more per task.
That does not make Le Chonk a bad deal. Its price will vary by provider and may change after the preview. And value per dollar is not the only reason to pick a model. Control, privacy and the ability to run it yourself are things this table cannot measure.
But it does mean one thing: “cheap per token” is not the same as “cheap per result”. Always compare the cost per finished task.
8. Who Should Actually Care About Le Chonk?
Not everybody needs to drop what they are doing for a trillion-parameter model. Here is a simple guide.
| If you are... | Should you care? | Why |
|---|---|---|
| A bank, hospital or government office with strict data rules | Yes, a lot | Self-hosting keeps your data in your own walls |
| A European company that wants AI independence | Yes | A serious open option built in Europe |
| A startup running AI features at huge volume | Maybe | Test the cost per finished task, not the price per token |
| A security team | Watch closely | Strong claims, but they still need independent proof |
| A developer building coding agents | Not yet | Opus 5.5 leads clearly on agentic coding |
| A blogger or everyday chatbot user | Not much | The apps you already use do the job |
You Should Care If Control Matters More Than Rank
If your biggest worry is “where does my data go?”, Le Chonk is the most interesting release of the month. Open weights let you run the model under your own rules, and nobody can change your price next quarter.
The money behind it is worth noticing too. AI is an expensive race, and the numbers keep growing. I looked at what these companies are spending and promising in The AI Bubble: OpenAI, Anthropic and SpaceXAI by the Numbers.
Security Teams Should Wait for Proof
Mistral’s cyber results are its biggest selling point. But the numbers are the company’s own. One reviewer could not verify the 82 percent cyber score that Mistral claims, and called the model not a download-and-run product. Wait for independent tests before you bet your defence on it.
Everyone Else Can Skip the Hype
If you use ChatGPT or Claude to write emails, plan trips or study, you do not need Le Chonk. Pick the app you like and move on. This model is mostly for builders and organizations.
9. What to Watch on October 27?
The weights are due on October 27. Here is the short checklist I will be watching, and you can too.
- The license. This is the big one. Mistral Large 3 used Apache 2.0, but Mistral Large 2 had a research license that restricted commercial use. The license for Large 4 has not been announced. Everything about business use depends on this text.
- The hardware math. Mistral has not published its own hardware needs, so every sizing guide is planning math until real weights exist. Running big models also eats a lot of power, and I explained how that reaches your own bills in AI and Your Electricity Bill: How Data Center Power Demand Hits Households.
- Independent coding and security tests. Once the weights are public, researchers can test the model without Mistral in the room.
- The price after the preview. The launch discount will end.
- The exact specs. Check the model card for the active parameter count and the real context window before you trust any number, mine included.
I will update this post after the weights land, so come back after October 27.
10. Frequently Asked Questions
Is Mistral Large 4 better than Claude Opus 5.5?
No. On the independent index, Opus 5.5 scores 57.6 and Le Chonk scores about 38. Le Chonk wins on price per token and on control, not on raw intelligence.
Is Mistral Large 4 better than GPT-6 Astra?
Not overall. Astra scores about 53 against 38. Mistral claims a small edge on one visual grounding test, but that is a single result.
Is Le Chonk really open source?
It is open weights, which is not the same thing. The license has not been announced, so do not assume you can use it for anything.
Can I run Le Chonk on my own computer?
Not on a normal computer. Even the smallest estimates put the weights above what any single GPU can hold. It needs a multi-GPU server.
Is it really cheaper than Claude and GPT?
Per token, yes by a wide margin. Per finished task, it depends. Le Chonk talks a lot, and Claude Opus 5.5 at medium effort gave better value in my math from Part 3.
Is Mistral Large 4 safe?
That question is not settled. Mistral is red-teaming it with partners, and independent testing will tell us more after the weights are public.
11. Final Verdict: A Big Orange Cat With Real Claws
Here is where I land.
Claude Opus 5.5 is the smartest model of the three, and its effort dial gives you good value when you pick the right setting.
GPT-6 Astra is strong and a bit pricey, and it makes the most sense for light, quick jobs.
Mistral Large 4 is not the smartest, and it was never going to be. But look at how far it came. Mistral Large 3 scored 9 on the same index, and Large 4 scores 38. That is a giant leap for a European lab, and it comes with open weights.
So who wins? It depends on what you value.
- Want the best brain? Claude Opus 5.5.
- Want a good balance on light jobs? GPT-6 Astra.
- Want control, privacy and a European option? Le Chonk.
The funny name hides a serious message. Open models are getting closer to the closed giants, and that changes the whole race.
What do you think? Will you try Le Chonk after the weights land? Tell me in the comments, and share this post with a friend who keeps asking which AI to use.
Keep reading: