Mistral Is Back: Le Chonk Beats GPT-6 Astra and Claude

Mistral has been quiet at the top end for a while. Mistral Large 3 shipped in December 2025, and since then the open-weight frontier conversation has focused on Kimi, DeepSeek, GLM, and Qwen. On October 6, 2026, Mistral responded with Mistral Large 4, a 1-trillion-parameter model the team calls “Le Chonk.”

Try Apidog today

The headline number is a cyber benchmark where Large 4 scores 82% while Claude Opus 5.5 and GPT-6 Astra score near zero. That result is on Mistral’s launch page, but it needs context: the closed models refuse the task. This guide covers the specs, pricing, trade-offs, and a practical way to evaluate Large 4 against your own API workloads.

TL;DR

• What it is: A 1.05T-parameter mixture-of-experts model with 49B active parameters, a 1.6B vision encoder, text and image input, and a 1M-token context window.
• API model ID: mistral-large-4, available as a public preview. The alias is mistral-large-4-0.
• Weights: Expected by the end of October 2026; the license has not yet been published.
• Price: $1.36 input / $4.18 output per million tokens at list price. Preview pricing is $0.68 input / $2.09 output. Cached input is $0.07 per million tokens.
• Cyber benchmark: 82% on the AA Cyber Index vulnerability reproduction test. Opus 5.5 and Astra score near zero because they refuse the task.
• Main limitation: In human-rated coding, Claude Opus 5 scored 4.22 while Large 4 scored 3.74.

Why “Mistral is back” is a fair take

For most of 2026, the strongest open-weight models came from China. Mistral continued shipping Medium 3.5, Devstral 2, and Small 4, but did not have a model competing directly with frontier closed models in their strongest categories.

Large 4 changes that positioning. Mistral claims it “significantly outperforms any open-weight model developed in the US or Europe.” Its training and deployment story also targets teams with EU sovereignty requirements:

• Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs.
• Training ran in Mistral-operated European datacenters.
• Supports more than 160 languages, including every official EU language.
• Launched shortly after Mistral’s €3 billion Series D, which the company describes as the largest equity round raised by a European technology company.

Timing also matters. Large 4 arrived shortly after Reflection unveiled its open-weight Beam model, adding a third serious contender to an open-model race previously dominated by US and Chinese efforts.

The cyber headline: what the 82% result actually means

The result being shared most widely comes from the Artificial Analysis Cyber Index. One test asks a model to reproduce a real vulnerability in open-source software and then patch it. Mistral reports that Large 4 scores 82%, the highest score of any model in its comparison.

However, Mistral also notes that Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task.

Read the result this way:

Claim	What it actually means
“Large 4 beats Astra and Opus 5.5 at cyber”	On this test, the closed models decline to answer. Large 4 answers and usually succeeds.
“Large 4 is the best hacking model”	Not established. A refusal is a policy decision, not a capability ceiling.
“Large 4 is unsafe”	Also not established. Mistral reports that its average refusal rate on cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm is higher than every other open-weight model.

There is still meaningful capability behind the headline. Mistral reports that Large 4:

• Solves 93% of Cybench, a set of 40 capture-the-flag exercises.
• Resists 93.3% of attacks on the B3 Agent Security Benchmark.

For security teams, the practical use case is defensive work: reproduce and patch a CVE in code you own, review a staging environment, or use the model as a second reviewer for security findings. For API teams, self-hosted availability means those workflows can eventually run under your own infrastructure and policy controls.

Where Large 4 beats GPT-6 Astra outside cyber

The cyber result is attention-grabbing, but the more useful comparisons are tasks where both models attempted the evaluation.

Benchmark	Mistral Large 4	GPT-6 Astra	Notes
Dense 200, visual grounding	42%	41%	A real but narrow one-point lead.
Finance Agent v2 (vals.ai)	Ahead	Behind	Mistral reports the ranking, not the scores.
Harvey Legal Agent Benchmark	Best open model	Listed	No direct head-to-head score published.

Against other open models, Mistral reports a wider lead:

Benchmark	Mistral Large 4	Who it beats
AutomationBench, 657 workflows	59.9%	Kimi K3, DeepSeek V4 Pro
AA-Briefcase, knowledge work	1,393 Elo	DeepSeek V4 Pro
DeepSWE v1.1, coding	61.7%	Leads open-weight models
SWE-Atlas-QnA	59.4%	—
Terminal-Bench 4.0	28.3%	—

These are Mistral-reported preview-model results. No independent lab has re-run them yet, and Euronews reports that reinforcement learning was still wrapping up at launch. Use the numbers as a strong signal, then validate the model on your own workload.

Where Large 4 still loses

Mistral published a comparison where a closed model clearly leads. In a Surge AI human evaluation of coding quality across five models, Large 4 Preview ranked second:

Model	Human coding score, out of 5
Claude Opus 5	4.22
Mistral Large 4 Preview	3.74
GLM-5.3	3.60
Kimi K3	3.59
GLM-5.2	3.40

That result puts Large 4 ahead of the compared Chinese open models, but roughly half a point behind Claude Opus 5.

Mistral’s framing—that Large 4 is “closing the gap” with frontier models in coding—is accurate. If coding quality is your primary metric, do not replace Claude based on benchmark charts alone. Run your repository tasks, pull-request reviews, test-generation prompts, and debugging workflows through both models.

Also note what is missing: the launch materials do not include a comparison against Claude Fable 5.1 or GPT-6 Sol. Treat any claim that Large 4 beats either model as unproven until a benchmark is published.

Specs at a glance

Spec	Mistral Large 4
API model ID	mistral-large-4
Alias	mistral-large-4-0
Version	26.10, public preview
Architecture	Mixture-of-experts, hybrid instruct + reasoning
Parameters	1.05T total, 49B active, 1.6B vision encoder
Context window	1M tokens
Input	Text and images
Reasoning	reasoning_effort: "high" or "none"
Tools	Function calling, structured outputs, built-in agent tools
Endpoints	/v1/chat/completions, /v1/conversations, /v1/agents, /v1/batch
Weights	Open weights expected by the end of October 2026; license not yet published

Call Large 4 from the API

Use the standard Mistral chat completions endpoint:

curl https://api.mistral.ai/v1/chat/completions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-large-4",
    "messages": [
      {
        "role": "user",
        "content": "Review this API error response and suggest a safer error-handling strategy."
      }
    ],
    "reasoning_effort": "high"
  }'

For fast, lower-cost requests where extended reasoning is unnecessary, set:

{
  "reasoning_effort": "none"
}

When evaluating the model, record these fields from the response:

• usage token counts for input and output cost calculations.
• Latency for comparable prompts.
• Structured-output validity when returning JSON.
• Functional correctness for your API, coding, or automation task.
• Refusal behavior for legitimate security and support workflows.

Price: a low-cost frontier-class API option

Mistral’s pricing page lists Large 4 at half its list price during the preview:

Model	Input / 1M tokens	Output / 1M tokens	Open weights
Mistral Large 4, preview price	$0.68	$2.09	Yes, late October
Mistral Large 4, list price	$1.36	$4.18	Yes, late October
Claude Opus 5.5	$4.00	$20.00	No
GPT-6 Astra	$10.00	$50.00	No

At preview pricing, Large 4 output is roughly one twenty-fourth of Astra’s output price. At list price, it is still about 12x cheaper on output.

Cached input costs $0.07 per million tokens. That can materially reduce spend for agent workflows that resend the same system prompt, OpenAPI definitions, tool schemas, and conversation context on every request.

Mistral has not said when the preview discount ends, so use list pricing for long-term production budgeting.

Should you switch?

Try Large 4 now if:

• You run agentic workflows or knowledge-work automation and pay frontier-model rates.
• You need multilingual output across EU languages.
• You need EU-hosted inference for compliance or sovereignty requirements.
• You do defensive security work and encounter refusals on legitimate vulnerability reproduction.
• You want to evaluate a frontier-class model that Mistral says will be self-hostable within weeks.

Wait before switching production traffic if:

• Coding quality is your most important metric. Claude still leads on the published human-rated coding result.
• You need independently reproducible benchmark numbers for an engineering or procurement review.
• You currently run Large 3 in production. Large 4 is still a preview and may change before a stable versioned release.

How to test Large 4 against your own APIs in Apidog

Benchmarks measure someone else’s tasks. The practical decision is whether Large 4 improves quality, latency, cost, or tool use for your prompts and APIs.

Apidog lets you create comparable API requests and turn them into repeatable regression tests.

1. Save the base request. Create a request for POST https://api.mistral.ai/v1/chat/completions. Store the API key in an environment variable and configure:

   Authorization: Bearer {{MISTRAL_API_KEY}}
   Content-Type: application/json

1. Create a baseline request body. Start with a prompt from a real workflow, such as API documentation generation, incident analysis, test-case generation, or structured extraction.

   {
     "model": "mistral-large-4",
     "messages": [
       {
         "role": "user",
         "content": "Return a JSON object containing API risks, severity, and recommended mitigations."
       }
     ],
     "reasoning_effort": "high"
   }

1. Clone the request for each model. Duplicate the request and change only the model field. Keep the prompt, temperature, output schema, and tool definitions identical. This makes latency, output quality, and token usage comparable.

2. Add assertions. Validate:

   • A 200 status code.
   • A non-empty answer.
   • Required fields in the response.
   • JSON Schema validity for structured output.
   • Expected content for known regression prompts.

3. Compare cost using the response usage block. Multiply input and output token counts by each model’s published rates. Do not compare only request-level cost; agent workflows can amplify output and repeated-context costs.

4. Rerun on every model update. Group requests into a test scenario and rerun them when Mistral updates the preview or releases weights. This helps catch quality drift before users do.

Two Large 4 capabilities map directly to API workflows:

• Use its security strengths as a second reviewer for API security testing, such as reproducing an authorization bypass in an API you own and control.
• Use function calling and structured outputs to turn an OpenAPI specification designed in Apidog into tool definitions for an agent.

For a full implementation walkthrough covering API keys, reasoning chunks, image input, function calling, and cost calculations, see the Mistral Large 4 API guide. If you are upgrading from an earlier Mistral model, the Mistral AI API basics and Mistral Medium 3.5 API guide still apply: same base URL, same authentication, new model ID.

FAQ

Is Mistral Large 4 open source?

It is open-weight, not necessarily open source. Mistral says weights will ship by the end of October 2026. The license has not yet been published, so do not assume it matches Large 3’s Apache 2.0 license.

Does Mistral Large 4 really beat GPT-6 Astra?

According to Mistral’s reported numbers, it leads Astra on visual grounding, 42% versus 41%, and on Finance Agent v2. On the cyber vulnerability test, Astra scores near zero because it refuses the task, so that is not a like-for-like capability comparison.

Does Large 4 beat Claude?

Not on the published human coding evaluation. Claude Opus 5 scores 4.22 while Large 4 Preview scores 3.74. Claude Opus 5.5 also refuses the cyber reproduction task, which explains its near-zero score on that evaluation.

What does “Le Chonk” mean?

It is Mistral’s unofficial nickname for the model and a joke about its size. The official name is Mistral Large 4, or ML4.

Can I use it for free?

Mistral’s free plan includes $10 per month in API credits and access to its latest models in Le Chat and Vibe. At preview pricing, $10 buys roughly 4.8 million output tokens on Large 4.

Bottom line

Mistral Large 4 puts Mistral back in the frontier-model conversation. By Mistral’s numbers, it is the strongest open-weight model from outside China, offers low API pricing, and is expected to become self-hostable within weeks.

The cyber claim is real but needs the refusal context. The more balanced conclusion is that Large 4 is highly competitive for automation, multilingual, visual, and defensive-security tasks, while Claude remains ahead on the published human-rated coding evaluation.

Test it against your real prompts, API schemas, tool calls, and regression cases while preview pricing is available. Then make a production decision after the weights ship and independent benchmarks arrive.