On October 6, Mistral AI put a public preview of Mistral Large 4 — internally and now publicly nicknamed “Le Chonk” — live on its API. It’s a 1.05-trillion-parameter, natively multimodal mixture-of-experts model, with 49 billion of those parameters active on any given request, and Mistral is benchmarking it directly against the current open-weight leaders out of China and the US. Open weights are due by the end of October. That’s the headline. The part nobody’s reconciled yet is smaller and more annoying: Mistral’s own announcement lists one price, and the live product page bills you a different one. Exactly half, to be specific.

Quick Summary: What Happened

Detail	Info
Announced	October 6, 2026, via Mistral’s official blog
Model	1.05T parameters, 49B active (MoE), natively multimodal, text output only
Trained on	3,800 Nvidia Grace Blackwell GPUs, Mistral’s own European datacenters
API preview	Live now; open weights expected around October 27
Price sheet says	$1.36 / million input tokens, $4.18 / million output
Live product page says	$0.68 / million input, $2.09 / million output — exactly half
Explanation given	None from Mistral directly; third-party trackers describe it as an unlabeled or quietly-labeled launch discount

Bottom line: Le Chonk is a legitimate flagship, and the European-sovereignty pitch is real. But if you’re an enterprise buyer trying to budget against Mistral’s own numbers right now, you’re budgeting against a coin flip.

What Le Chonk Actually Is

Start with the model, because it’s a genuinely large release. Mistral Large 4 is a 1-trillion-parameter mixture-of-experts model — MarkTechPost put the precise count at 1.05 trillion total parameters, with roughly 49 billion active per token. That’s the MoE trade-off in a sentence: you get the knowledge capacity of a trillion-parameter model while only paying the inference cost of a much smaller one. It’s natively multimodal on the input side (text and images), but answers in text only.

It was trained from scratch, over about two months, on 3,800 Nvidia Grace Blackwell GPUs — in Mistral’s own datacenters, in Europe, according to VentureBeat’s reporting on the launch. That detail isn’t filler. Mistral is explicitly selling this as a sovereign alternative — a frontier model that a European bank, hospital, or ministry can run without a single byte touching US or Chinese infrastructure. Training data spans more than 160 languages, including every official EU language. The pitch is coherent: if “AI sovereignty” is going to mean anything concrete in Europe, it needs a model that’s actually competitive, not just locally hosted.

The Benchmarks Mistral Is Leading With

Mistral isn’t benchmarking Le Chonk against its own last release. It’s benchmarking against the current open-weight frontier, full stop: DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3, and GLM-5.3. The headline numbers, straight from the announcement:

• 93% on Cybench — a benchmark built from roughly 40 exercises pulled from real security competitions
• 82% on CyberGym-E2E — a reproduce-and-patch style vulnerability test
• 61.7% on DeepSWE v1.1 — a software-engineering agent benchmark
• 49.8% on the Coding Agent Index — which Mistral says puts it ahead of DeepSeek V4 Pro and Qwen3.8 Max specifically

Cybersecurity is clearly where Mistral wants this model to live. That’s not an accident, either — it’s launching in preview specifically to developers, “cybersecurity leaders,” and government bodies before a wider release, which reads like a company trying to get its sharpest edge validated by the people who’d actually catch a weak claim.

Worth sitting with, though: Mistral’s own benchmarks are Mistral’s own benchmarks. Artificial Analysis, the independent evaluation firm, ran its own numbers and landed on an Intelligence Index score of 38 — “comparable to GPT-6 Luna (max) at 38 and DeepSeek V4.1 Flash (max) at 39,” and ranking 64th out of 225 models it tracks. Artificial Analysis’s framing was generous in one specific way — “the most intelligent AI model from outside the US and China” — and pointed in another: by its measurement, Le Chonk costs more than four times as much per completed task as similarly-capable open-weight competitors. That’s a materially different story than “leads the open-weight field,” and it’s the kind of gap between vendor claims and independent testing this site flags whenever it shows up. If you’ve been tracking how DeepSeek V4 and GLM-5 stack up against each other, Le Chonk slots into that conversation as a strong specialist, not a clean leader across the board.

The Pricing Problem Nobody’s Reconciled

Here’s where it gets genuinely confusing, and where most of the coverage from launch day missed each other entirely.

Mistral’s own announcement post, and the figures VentureBeat and MarkTechPost both cited at launch, list Mistral Large 4 at $1.36 per million input tokens and $4.18 per million output tokens. That’s the number that’s been repeated across most of the launch-day write-ups.

But pull up the live pricing on Mistral’s product page, or check it through a routing service like OpenRouter or Vercel’s AI Gateway, and the number is $0.68 input / $2.09 output — precisely half. One outlet, Forkast, covering the pricing angle from the “this undercuts everyone” direction, cited the $0.68/$2.09 figures as simply Mistral’s rate — no mention that a different number exists anywhere else, including in Mistral’s own announcement.

So which is it? The most coherent explanation going around — and it’s a plausible one, not a confirmed one — is that the lower figures are a launch-week discount: roughly a 50% markdown for the first two weeks of preview, after which pricing reverts to the $1.36/$4.18 list rate. If that’s right, the discount window lines up with an expiration around October 20. But Mistral hasn’t put that explanation in writing anywhere I can find, and “probably a two-week promo, going by how the math lines up” is not the same as Mistral confirming it. For a model being pitched partly on enterprise trust and partly on EU sovereignty, shipping with an ambiguous price sheet on day one is a strange way to start that conversation.

Why this matters more than a typical pricing typo

If you’re evaluating Le Chonk for a production workload, the difference between $1.36/$4.18 and $0.68/$2.09 isn’t rounding error — it roughly doubles or halves your projected spend depending on which number survives past the preview window. Enterprises comparing this against GPT-6.1 Sol’s pricing or any other frontier API need a number they can actually budget against, not two numbers and a guess about which one is real.

Who Should Actually Care About This

European enterprises and public-sector buyers chasing AI sovereignty are the obvious audience, and the infrastructure story here is genuinely strong — Mistral trained this in its own datacenters, under European law, independent of US or Chinese cloud providers. That’s a real differentiator if Apple’s EU regulatory fights over Siri have taught you anything about how fast “where’s the data actually processed” becomes a boardroom question in Europe.

Security teams evaluating agentic tooling should pay attention to the Cybench and CyberGym numbers specifically — they’re the strongest, most differentiated results in the release, and the kind of thing worth testing directly rather than taking on faith.

Anyone doing procurement math should wait for Mistral to publish a single, stable price before locking in a budget line. Build your cost model around the $1.36/$4.18 list figures until Mistral says otherwise in writing — modeling the discounted rate as your steady-state cost is how budgets go sideways three weeks into a contract.

If you’re weighing Mistral generally against building on your own infrastructure, our breakdown of Mistral Forge covers the self-hosting and customization side of that decision, and Mistral’s Voxtral TTS release is a useful reference point for how the company has handled open-weight pricing and positioning in the past.

The Bigger Picture

Open weights land around October 27, which is the date that actually settles most of the “is this real” questions. A vendor-run cybersecurity benchmark is one thing; a trillion-parameter model that security researchers can pull apart, run locally, and stress-test themselves is a different kind of proof. Until then, Le Chonk is a strong claim backed by Mistral’s own numbers, a more modest but still respectable independent score from Artificial Analysis, and a price tag that currently depends on which page you’re looking at.

Our Take

We think the model is the easy part of this story. Mistral has shipped credible open-weight models before, the Grace Blackwell training run is a real capital commitment, and the EU-sovereignty pitch is more substantive than most “sovereign AI” marketing because the infrastructure claim is specific and checkable. What’s harder to wave off is a pricing page that contradicts the company’s own launch announcement, with zero public acknowledgment that two numbers exist. That’s not a scandal. It’s sloppy, in a way that matters more for Mistral than it would for a smaller vendor, because “trust us with your sovereign data” and “we can’t keep our own price sheet consistent” are an uncomfortable pair to ship in the same week.

Frequently Asked Questions

What is Mistral Large 4, and why is it called “Le Chonk”?

Mistral Large 4 is a 1.05-trillion-parameter, open-weight, natively multimodal mixture-of-experts model from Mistral AI, announced October 6, 2026. “Le Chonk” is Mistral’s own nickname for the release, referencing its size — 49 billion of its parameters are active on any given request, with the rest distributed across its mixture-of-experts architecture.

When do the open weights release?

Mistral says weights will be published by the end of October 2026, with multiple outlets citing an expected date around October 27, following a roughly three-week preview period with developers, security researchers, and government partners.

What’s the real price of Mistral Large 4?

As of this writing, that’s unresolved. Mistral’s own launch announcement and most launch-day coverage cite $1.36 per million input tokens and $4.18 per million output tokens. The live API and third-party routers like OpenRouter currently bill exactly half that — $0.68 input / $2.09 output. The most plausible explanation is an unannounced launch-week discount, but Mistral hasn’t confirmed this in writing.

How does Le Chonk compare to DeepSeek, Qwen, Kimi, and GLM-5.3?

Mistral’s own benchmarks claim leadership over DeepSeek V4 Pro and Qwen3.8 Max on its Coding Agent Index, and strong results against Kimi K3 and GLM-5.3 on cybersecurity-focused tests. Independent testing from Artificial Analysis placed it closer to parity with comparable models — an Intelligence Index score of 38, similar to GPT-6 Luna (max) and DeepSeek V4.1 Flash (max) — while costing roughly four times more per task than similarly-capable open-weight alternatives.

Is Mistral Large 4 actually trained and hosted in Europe?

Yes, according to Mistral and corroborating reporting. The model was trained on 3,800 Nvidia Grace Blackwell GPUs inside Mistral’s own European datacenters, and the preview is served from that same infrastructure, operating independently of US and Chinese cloud providers and under European law.

Should enterprise buyers commit to Mistral Large 4 pricing right now?

Not without a confirmed number from Mistral. Budget against the higher, officially-announced $1.36/$4.18 figures rather than the current discounted rate, since the lower price is most likely tied to a temporary launch promotion that could expire within weeks of the October 6 preview.

---

Last updated: October 7, 2026. Sources: Mistral AI — Introducing Mistral Large 4 · VentureBeat — Mistral debuts Large 4 “Le Chonk” · MarkTechPost — Mistral AI Releases Mistral Large 4 · Forkast — Mistral Large 4 Doesn’t Just Compete With Chinese Open-Weights · Unite.AI — Benchmark Firm Gives Mistral Large 4 Preview a 38 Intelligence Score · OpenRouter — Mistral Large 4 pricing.

Related reading: Mistral Forge Review · Mistral Voxtral TTS: Is Open-Source Voice AI Ready? · DeepSeek V4 Review · GLM-5 Review · Apple’s Siri AI Blocked in EU