Mistral has launched the public preview of Mistral Large 4, nicknamed "Le Chonk": a 1-trillion-parameter, natively multimodal model with 49 billion active parameters, trained from scratch on 3, 800 NVIDIA Grace Blackwell GPUs in the company's own European datacenters. The preview API is live now at $1.36 per million input tokens and $4.18 per million output tokens. We verified current API pricing across eight frontier models and calculated the blended cost per million tokens for each. Large 4 lands at about $2.06, roughly one-tenth the cost of GPT-6 Astra and half the cost of Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon.
What launched
On October 6, 2026, Mistral announced Mistral Large 4 (ML4), "unofficially ML4, very officially: le Chonk, " as a public preview through its API on Mistral Studio. The company calls it its largest and most capable model to date. The weights are promised by the end of October, with Reuters reporting October 27 as the target date, after a red-teaming period with cybersecurity leaders, vetted partners and state authorities.
The verified specifications, from Mistral's own announcement:
- 1 trillion total parameters, 49 billion active per token (sparse mixture-of-experts)
- Natively multimodal: text and image input, text output
- Trained from scratch on 3, 800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters
- Training data spanning 160+ languages, including every official EU language
- Context window of about 1 million tokens, per the model documentation
On benchmarks, Mistral reports 93% on Cybench and 82% on a vulnerability reproduction and patching test, 61.7% on DeepSWE v1.1, a combined Coding Agent Index score of 49.8% (ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max), 59.9% on AutomationBench, and 42% on Dense 200 visual grounding, narrowly ahead of GPT-6 Astra at 41%. Notably, Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on the vulnerability reproduction test because they refuse the task, a deliberate positioning of open weights for defensive security work.
The launch is the first milestone funded by Mistral's €3 billion Series D, which the company describes as the largest equity round ever raised by a European technology company.
The ByteHerald price check
Launch announcements rarely tell you what a model costs to actually use. So we pulled the current per-token API rates for eight frontier models and computed a blended cost per million tokens at a 3:1 input-to-output ratio, a common mix for agent and assistant workloads. Rates were checked on October 7, 2026, against provider pricing pages wherever possible.
| Model | Input / 1M | Output / 1M | Blended / 1M |
|---|---|---|---|
| GPT-6 Astra (OpenAI) | $10.00 | $50.00 | $20.00 |
| Claude Opus 5.5 (Anthropic) | $4.00 | $20.00 | $8.00 |
| Gemini 4 Argon (Google, introductory) | $2.00 | $10.00 | $4.00 |
| Claude Sonnet 5.5 (Anthropic) | $2.00 | $10.00 | $4.00 |
| GPT-6.1 Sol (OpenAI) | $2.00 | $10.00 | $4.00 |
| Mistral Large 4 (preview) | $1.36 | $4.18 | $2.06 |
| DeepSeek V4 Pro (list rate) | $0.99 | $2.97 | $1.49 |
| Mistral Large 3 | $0.50 | $1.50 | $0.75 |
Three findings stand out. First, Mistral Large 4 undercuts every Western frontier model on price by a wide margin: about one-tenth of GPT-6 Astra, one-quarter of Claude Opus 5.5, and half of the $4.00 cluster formed by Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon. Second, the MoE architecture is doing real economic work here: only 4.9% of the trillion parameters activate per token, so inference prices like a mid-tier model rather than a trillion-parameter one. Third, the price is not the whole story: DeepSeek V4 Pro is cheaper still at about $1.49 blended, and Mistral's own Large 3 costs less than half of Large 4 at $0.75, so buyers are paying a clear premium for the new model's cyber and agentic capabilities.
Two caveats matter. The $1.36/$4.18 preview rate is introductory: reporting indicates a 50% launch discount for the first two weeks, and preview pricing generally does not survive contact with general availability. And "open-weight" is currently a promise, not a fact: until the weights drop at the end of October, the only way to run Large 4 is Mistral's hosted API, and the license for the weights has not been announced.
Our verdict
Mistral Large 4 is the most interesting open-weight launch of the year for a specific buyer: European enterprises and security teams that need frontier-class cyber and agentic capability without US-vendor lock-in, and that can wait a few weeks for weights to self-host. On price alone it is genuinely disruptive among Western labs. But anyone choosing on cost per token today should also price DeepSeek V4 Pro, and anyone choosing on capability should wait for independent benchmark runs before treating Mistral's first-party numbers as final. We will re-run this price check when the weights and final pricing land.
Methodology
Input and output rates were taken from OpenAI's official API pricing page (platform.openai.com/docs/pricing) for GPT-6 Astra, GPT-6.1 Sol; from Anthropic's official pricing page as cited in current technical coverage for Claude Opus 5.5 and Sonnet 5.5; from launch reporting for Mistral Large 4 ($1.36/$4.18) and Gemini 4 Argon introductory pricing ($2/$10); and from DeepSeek's official pricing documentation as cited in current coverage for V4 Pro list rates ($0.99/$2.97). Blended cost assumes 750, 000 input tokens and 250, 000 output tokens per million. Prices change frequently; check the provider's live pricing page before budgeting. No provider paid for or reviewed this analysis.
Related reading
OpenAI's AI Wrote 722 Math Manuscripts. We Read the Repo: 3 Withdrawn in 24 Hours
Our data-driven breakdown of what OpenAI's unreleased frontier model actually produced: 372 result families, 300 Lean-formalized, 3 withdrawn in a day.
By ByteHerald Staff. Published October 7, 2026.
Comments
Post a Comment