Mistral Large 4 (Le Chonk): Pricing, Benchmarks, and Open Weights

The model, which Mistral AI nicknames Le Chonk, is the company's largest: a natively multimodal mixture-of-experts with about 1 trillion total parameters, announced on October 6, 2026. Mistral's pitch has three parts: the strongest open-weight model built in the US or Europe, state-of-the-art among open models on cybersecurity, finance, and law, and a model trained and served on Mistral's own European infrastructure. Below: price and specs as documented, Mistral's benchmark claims with their footnotes, what independent testing shows, what European processing requires, and what is not yet known, starting with the license. No hands-on test is included: the article compares published claims with independent results.

Mistral Large 4 at a glance

• Status: public preview on the Mistral APIAPIAn API (application programming interface) is a defined way for one piece of software to request data or actions from another. AI companies offer APIs so developers can build their models into other products.Read the full definition from October 6, 2026; open weights promised by the end of October
• Architecture: mixture-of-experts, natively multimodal with a 1.6 billion-parameter vision encoder; the launch post says 1 trillion total and 49 billion active parameters, the docs model page says 1.05 trillion and 52 billion
• Price: $0.68 input, $0.07 cached input, $2.09 output per million tokens for two weeks from launch; regular price listed as $1.36 and $4.18
• Context: 1 million tokens in Mistral's documentation; Artificial Analysis lists 524K
• API model name: mistral-large-4
• Languages: more than 160, including every official EU language
• Training: from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters
• Not yet announced: the license for the open weights

What is new in Mistral Large 4?

1. Mistral's largest model, with two sets of numbers

Mistral Large 4 is the first Mistral model at the trillion-parameter scale. The parameter figures differ by source: the X announcement says 1 trillion total and 49 billion active, the docs model page says 1.05 trillion and 52 billion. Both are Mistral's; the company has not explained the difference. Mistral also says the reinforcement learning run behind the preview is still in progress, so the model served this week may differ from the one whose weights ship.

2. Price: half off for two weeks, then above Mistral Large 3

Source: Mistral's model page, pricing page, and changelog, October 7, 2026. The pricing page shows $0.68 and $2.09 as the Standard sale price with $1.36 and $4.18 as the original price; the changelog entry for October 6 says "Launch pricing: 50% off for 2 weeks." The announcement and the X post quote the regular price. Batch processing is a separate tier at a further 50% off Standard, and cached input cuts repeated prompt cost by up to 90%.

Against Mistral Large 3, the launch price is about 1.4 times per token; once the discount ends, the regular price is about 2.7 times. Mistral has not said whether the regular price will change when the weights ship.

3. Benchmark claims, with the footnotes

Mistral publishes selected results rather than one table, and several come from third parties it engaged.

Mistral explains the cyber result specifically: several closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the reproduce-and-patch test because they refuse it, while proving a flaw is real is the first step of defending software. That argument applies to the cyber test; Mistral does not make it for finance, law, or grounding.

What independent testing shows

Artificial Analysis had scored the preview by October 7, 2026, on two fronts.

General index. The preview scores 38 on the firm's ten-benchmark Intelligence Index, 64th of 225 models, generating 200 million output tokens across the suite against an 81 million median, at 116 tokens per second. The firm's cost figure of $1.13 per index task is computed at the regular $1.36 and $4.18 rates, not the launch price, so it overstates the cost of the model during the discount; the ranking does not depend on price.

Source: Artificial Analysis model pages, index v4.3.2, October 7, 2026. On this composite, Mistral's statement that the model is competitive with the strongest open-sourceOpen sourceOpen source software has public source code that anyone can read, use, and modify under the terms of its license. Many AI tools and agent frameworks are released this way.Read the full definition models globally holds only partly: it sits 6 to 8 points below the Chinese leaders and ahead of DeepSeek V4 Pro. The claim to be the best open-weight model from the US or Europe cannot be checked against a published list; Artificial Analysis's open-weight leaderboard is led by Chinese models.

Cybersecurity. On CyberGym-E2E-AA, Artificial Analysis's independent implementation of a defensive benchmark in which an agent must find a real memory-safety vulnerability in an open-source C or C++ project, reproduce it with a proof-of-concept input, and patch it without breaking the project's tests, Mistral Large 4 Preview ranks first at 81.7% pass@1, ahead of MiMo-V2.6-Pro at 78.6% and GPT-6 Luna at 77.9%. The leaderboard also records tasks declined on safety grounds, which is where refusing models lose points. This confirms Mistral's reproduce-and-patch claim on one test. It does not by itself confirm the top-five placement on the composite Cyber Index, which has not been independently confirmed.

European processing: what it requires

Mistral's announcement describes a European deployment operated end to end under European law, and the model was trained and the preview is served from Mistral's European datacenters. For a customer, the documentation adds conditions. The global endpoint, api.mistral.ai, does not commit to a processing location. EU processing requires the regional endpoint, api.eu.mistral.ai, which costs 1.1 times the Standard price, serves only the models hosted in that region, supports function calling as its only tool, and does not offer stateful features such as Agents, Batch, or the Files API. Account configuration, keys, billing, and usage analytics may still be handled outside the chosen region, and zero data retention is a separate control. Check that mistral-large-4 is listed on the EU endpoint before routing production traffic to it.

Mistral frames the release as the first milestone funded by its 3 billion euro Series D and says more compute is coming online in its own datacenters.

Cyber access during preview

During the preview, cybersecurity leaders, vetted partners, and state authorities get the same model with reduced moderation and expanded cyber capabilities; everyone else gets the moderated API, which Mistral says refuses malicious cyber prompts more often than other open models. Once the weights are released, organizations will be able to run the model on private cloud or on-premise under their own policies.

Open weights and the license question

Mistral says the weights will be released by the end of October 2026, with architecture details, more benchmarks, and the post-training methodology. The license is not announced. The docs overview tags Mistral Large 4 as OPEN where Mistral Large 3 is tagged Apache 2.0 and Mistral Medium 3.5 Modified MIT, and Mistral's pricing FAQ notes that some open-weight models are Apache 2.0 for research and individual use while commercial deployment requires a Mistral license. Until the license is published, plan self-hosting as a possibility, not a commitment.

Who should use Mistral Large 4

• Security teams and incident responders: the clearest fit, now with one independent result behind it. Test the preview API on vulnerability reproduction, detection rules, and malware triage; apply for partner access if you need the less-moderated version.
• European organizations with data-location requirements: evaluate through the EU regional endpoint at the 10% premium, confirm the model is listed there, and note the missing stateful features. The license decides whether self-hosting is an option.
• Developers choosing an open-weight model for agents or coding: compare on your tasks against MiMo-V2.6-Pro, GLM-5.3, and Kimi K3, which score higher on the independent index, some at lower cost per task.
• Teams on frontier closed models for terminal work and long office tasks: the published gap is large; Mistral's case here is price and control, not capability. For the closed model Mistral most often measures itself against, see GPT-6 Astra: What OpenAI Released and What It Means.

What is not yet known

What this means for small business owners

Launches like Mistral Large 4 matter most to developers, security teams, and companies with data-location rules. For the owner of a small service business, the useful question is simpler: which everyday tasks can be handed off, and how is the result checked before it goes out. Empacer is a voice-first AI business assistant for small service business owners. You say what needs doing, review the finished work, and decide whether it goes out.