Mistral Large 4: Le Chonk previews enterprise agents before the weights arrive
By AgentRiot Editorial
Mistral's new flagship has credible coding, cyber and document results. Its public preview offers discounted tokens, but independent tests show why agent buyers should watch the whole task bill.

Mistral Large 4 is available to try through an API, with downloadable weights promised later in October. Mistral launched the public preview on October 6 (opens in new tab), calling the model Le Chonk and pitching it for agents that write code, investigate security flaws and produce professional documents.
Independent evaluations put the preview alongside GPT-6 Luna on one general intelligence index, with a strong vulnerability-reproduction and patching result. They also expose a commercial catch: inexpensive tokens can still produce an expensive agent run when the model generates a lot of them.
API now, weights later
Mistral's launch announcement (opens in new tab) promises weights by the end of October. CNET's interview with chief scientist Guillaume Lample and VP of Science Pierre Stock (opens in new tab) reports an October 27 rollout, a narrower date than Mistral's published commitment. The current model card (opens in new tab) marks both weights and license “Coming soon.” Self-hosting and license decisions must wait for the release.
Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacenters, which also serve the preview. For enterprises seeking eventual control over deployment, that planned weight release is more consequential than the nickname.
The published specifications do not yet agree. The announcement describes one trillion total parameters with 49 billion active; the model card lists 1.05 trillion total, 52 billion active and a 1.6-billion-parameter vision encoder. The card advertises a one-million-token context, while Artificial Analysis (opens in new tab) and Vals AI (opens in new tab) list 512k; AA's model page displays 524k. Confirm the endpoint's limit before designing a workflow around one million tokens.
Coding and documents: useful, uneven gains
Artificial Analysis gives the preview 38 on Intelligence Index v4.3.2 (opens in new tab). Its launch analysis (opens in new tab) places GPT-6 Luna at 38 and DeepSeek V4.1 Flash at 39, both at maximum reasoning. That is a competitive result, not a claim that Le Chonk beats every frontier model.
Mistral reports (opens in new tab) 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index score of 49.8%. Vals reports 22.73% on Terminal-Bench 4.0 (opens in new tab), a lower result whose cause these pages do not explain.
A blind coding-quality evaluation Mistral ran with Surge AI placed the preview second among five models. Professional annotators gave it 3.74 out of five, below Claude Opus 5's 4.22 and above Kimi K3's 3.59. The preview is competitive here, but Opus retains a clear lead in the published ratings.
For business agents, Mistral reports 59.9% on AutomationBench's 657 workflows and 1,393 Elo on AA-Briefcase (opens in new tab). These tasks involve using applications and producing deliverables, rather than answering isolated chat questions.
The multimodal results are mixed too. Mistral's Dense 200 visual-grounding result is 42%, narrowly above GPT-6 Astra's 41% in its comparison. Independently, AA measures 19% on GDP.pdf, tied with MiMo-V2.6-Pro but behind Kimi K3's 22%. Locating objects in an image and reasoning over a professional PDF are different jobs; neither result establishes blanket vision leadership.
Cyber strength, with a narrower legal claim
Le Chonk's strongest independent result is specific. AA reports 82% on CyberGym-E2E-AA (opens in new tab), ahead of MiMo-V2.6-Pro at 79% and GPT-6 Luna at 78%. The benchmark asks models to reproduce real software vulnerabilities and patch them. It is not an 82% production vulnerability-fix rate.
On AA's broader Cyber Index, the preview scores 50, tied with GLM-5.3-Flash and behind MiMo-V2.6-Pro's 56. Mistral separately reports solving 93% of Cybench's 40 security-competition tasks. Provider refusals can also affect cyber comparisons, so a low score does not always mean a model lacks the underlying capability.
Access has an important boundary: Mistral says vetted partners receive reduced moderation and expanded cyber capabilities during red-teaming (opens in new tab). That does not make the public preview unrestricted.
Vals AI's results (opens in new tab) give a less sweeping picture of professional work. The model scores 48.05% overall, ranking 32nd of 44 on the Vals Index. Finance Agent v2 reaches 54.68%, ranking 22nd of 75; Harvey's Legal Agent Benchmark reaches 15.83%, ranking sixth of 75. The legal rank is strong, but it is not a promise of reliable autonomous legal work. Vals lists high reasoning as its default, with settings potentially differing by benchmark.
Mistral also reports (opens in new tab) 93.3% attack resistance on B3 and 1.691 out of two on KORA. Those safety scores do not remove the need to restrict an agent's tools and review consequential actions.
Discounted tokens are not necessarily cheap tasks
As of October 7, Mistral's pricing table (opens in new tab) lists these rates per million tokens:
| Token type | Sale | Original |
|---|---|---|
| Input | $0.68 | $1.36 |
| Cached input | $0.07 | $0.14 |
| Output | $2.09 | $4.18 |
AA describes the discount as a two-week launch offer (opens in new tab). Mistral's table labels the rates “Sale” and “Original” without specifying an end date.
A synthetic workload with one million uncached input tokens and 250,000 output tokens costs $1.2025 at sale rates, versus $2.405 at original rates. That calculation excludes tools, caching, batch adjustments, tax and retries. It is a token budget, not the cost of completing an agent task.
AA's measured task economics are less flattering. Its Intelligence Index costs $1.13 per task at original prices, or $0.57 with the discount, compared with $0.25 for GLM-5.3-Flash and $0.27 for DeepSeek V4.1 Flash at maximum reasoning. Those are index-specific weighted costs, not universal enterprise tariffs.
AA records 200 million output tokens across the index, versus an 81-million-token median (opens in new tab). It also measures 116.1 output tokens per second. Fast generation helps; long reasoning traces and extra agent turns still consume time and tokens.
What to test in the preview
The model card (opens in new tab) lists function calling, structured outputs and document Q&A. The model accepts text and images and produces text, making repository investigations and document-heavy workflows sensible trial candidates.
For integration, Mistral's reasoning documentation (opens in new tab) explicitly supports mistral-large-4-0 with reasoning_effort="high" or "none". High reasoning returns a content list containing ThinkChunk and TextChunk; none returns a plain string. Multi-turn clients should preserve the full assistant message, including thinking chunks, in history rather than reconstructing only the visible answer.
Enterprise teams should compare completion quality and total token consumption on their own workflows, especially where tool calls have consequences. The preview is worth evaluating for cyber and document agents now. A self-hosting decision must wait for the actual weights, license and deployment requirements; a purchasing decision should depend on the cost of a reviewed, successful result, not the cheapest number in the token table.

