On October 6, 2026, at the AI Everything summit in Abu Dhabi, Mistral AI unveiled Mistral Large 4, its new flagship model, nicknamed internally "Le Chonk". Trained entirely in European data centers, the multimodal model claims prices well below those of its American rivals. For an SMB comparing its AI options, here is what this new Mistral model actually changes.
In brief
- On October 6, 2026, Mistral AI launched Mistral Large 4 in preview on its API (Mistral Studio), with open weights expected on Hugging Face by the end of October 2026.
- The model claims roughly 1 trillion parameters in total, with 49 billion active at inference time, according to Mistral AI's own announcement.
- It was trained on 3,800 Nvidia Grace Blackwell GPUs, in data centers that Mistral operates itself in Europe, under European law.
- Its preview price ($0.68 per million input tokens) is up to 3 times cheaper than the floor set by GPT-6.1 Sol and Gemini 4 Argon ($2), according to a Yahoo Finance analysis.
- It claims the highest score of any tested model on a cybersecurity benchmark (vulnerability reproduction), a point that matters for SMBs focused on compliance.
A model built for digital sovereignty
Mistral Large 4 is a mixture-of-experts (MoE) model, natively multimodal: it handles text and images through a 1.6-billion-parameter vision encoder. It was trained on data covering more than 160 languages, including every official EU language.
The distinctive point of this launch is not purely technical. Mistral stresses that the training and deployment infrastructure remains entirely operated by the company in Europe, under European law, independent of other major cloud providers. The model is designed to be self-hostable by organizations that want full control over their data, an argument that speaks directly to SMBs handling sensitive customer data or GDPR-regulated information.
A nuance worth knowing
The technical sheet published by Mistral on its own platform lists a 675-billion-parameter model (41 billion active), a different figure from the one announced on stage (1 trillion / 49 billion active). Several observers flagged this discrepancy. It changes nothing for an SMB's practical use, but it is worth noting for accuracy.
Pricing that shakes up the market
The most striking commercial argument of this launch is price. In preview, Mistral Large 4 costs $0.68 per million input tokens and $2.09 per million output tokens, a rate expected to rise to $1.36 and $4.18 once the preview period ends.
| Tier | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Preview (Mistral Studio) | $0.68 | $2.09 |
| Standard tier (API) | $1.36 | $4.18 |
| Cached input tokens | $0.14 | - |
According to a Yahoo Finance analysis, the price floor set by competing closed models (OpenAI's GPT-6.1 Sol, Google's Gemini 4 Argon) sits around $2 / $10 per million tokens, and the average for proprietary models reaches $6.03. Microsoft, for its part, prices its MAI-Thinking-1 model at $2 / $8.
For an SMB running high token volumes (automated customer support, document processing, continuous AI agents), this price gap translates directly into a smaller monthly bill.
Performance aimed at enterprise use cases
Mistral is not only targeting consumers. The French company directed this launch's messaging toward specific sectors: cybersecurity, finance, law, manufacturing, engineering, logistics, pharmaceuticals and the public sector.
On Harvey's Legal Agent Benchmark, Mistral claims the best performance among open-weight models. In finance, the company says it outperforms GPT-6-Astra on both tested cases. These figures come from Mistral AI's own tests and have not yet been verified by independent benchmarks: a useful starting point, not an absolute truth.
The availability roadmap
October 6, 2026
API preview opens
End of October 2026
Open weights release
What this actually changes for an SMB
Three practical consequences follow from this launch. First, an SMB already processing large volumes of text with a closed model (OpenAI, Google, Anthropic) now has an additional negotiating or switching argument: a European model, cheaper to run, available through an API compatible with market standards (function calling, agents, structured outputs).
Second, for SMBs under strict data-residency requirements (health, finance, public sector, sensitive data processing on behalf of clients), the promise of a model trained and operated entirely in Europe, under European law, addresses a real concern, without relying on a non-European cloud partner to host the model itself.
Third, the planned open weights release at the end of October 2026 opens up an additional option: an SMB with in-house technical skills (or an IT provider) could, if it chooses, host the model on its own infrastructure rather than depend on an external API, a relevant choice for data that must never leave the company's own systems.
Before Mistral Large 4
Limited choice between closed American models (high costs, non-EU hosting) and open-weight alternatives often weaker on business use cases (law, finance, cybersecurity).
With Mistral Large 4
A European model, cheaper to run, with competitive scores on business benchmarks, and a full self-hosting option once open weights ship.
With measured optimism, this launch confirms that a European player can compete on price and performance against American and Chinese giants, without sacrificing sovereignty guarantees. For an SMB, it is one more concrete option to evaluate, not an immediate revolution to adopt without prior testing.
FAQ
What is Mistral Large 4?
It is Mistral AI's new flagship model, announced on October 6, 2026. It is a multimodal mixture-of-experts model, trained in Europe, available in preview through the Mistral Studio API, with open weights planned for the end of October 2026.
Is Mistral Large 4 cheaper than ChatGPT or Gemini?
In preview, its rate ($0.68 per million input tokens) is below the $2 floor set by GPT-6.1 Sol and Gemini 4 Argon, according to a Yahoo Finance analysis. The rate will rise to $1.36 once the preview ends, still below the market average ($6.03).
Can an SMB self-host Mistral Large 4?
Yes, once the open weights are published on Hugging Face, expected by the end of October 2026. This requires in-house technical skills or an IT provider, but gives full control over hosting and the data processed.
Are the performance figures Mistral announced reliable?
They come from Mistral AI's own tests and have not yet been confirmed by independent benchmarks. They offer a useful signal, but an SMB should ideally test the model on its own use cases before switching.
To go further on AI sovereignty, read our guide Sovereign AI: Mistral Paves the Way for SMBs in 2026 and browse all our resources on AI in business.
Sources: Mistral AI, official announcement "Mistral Large 4" (October 6, 2026); Yahoo Finance / AI, "Mistral Large 4 Doesn't Just Compete With Chinese Open-Weights" (October 6, 2026); MarkTechPost, "Mistral AI Releases Mistral Large 4 (Le Chonk)" (October 6, 2026).
