• Mistral Large 4 is a mixture-of-experts model with 1 trillion total and 49 billion active parameters, in public preview from October 6, 2026, per Mistral AI.
• Mistral lists API pricing of $1.36 per million input tokens and $4.18 per million output tokens.
• Mistral's post says weights will arrive at the end of October, but neither it nor Tech.eu states a license, so the model is not yet usable as an open-weight option.
• Every benchmark figure, including 93% on Cybench and 61.7% on DeepSWE v1.1, comes from Mistral and has no independent reproduction yet.
• The decision for a buyer is cost per completed task against a hosted API, which launch benchmarks do not answer.

Mistral AI launched Mistral Large 4 in public preview on October 6, 2026. Tech.eu reports that the company says the model significantly outperforms any open-weight model developed in the US or Europe. The claim is the vendor's own, the weights are not yet public, and the benchmark scores come from Mistral's announcement. For a team deciding what to build on, the useful question is what is confirmed today, what it costs, and what must be verified first.

What Mistral announced

Mistral's announcement describes a hybrid instruct-and-reasoning mixture-of-experts model with 1 trillion total parameters and 49 billion active at inference. It is natively multimodal and covers more than 160 languages. Mistral says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its European datacenters. A public preview is live now. The post says the weights will arrive at the end of the month, in its words 'Weights drop end of this month'.

Benchmarks are claims until reproduced

Mistral reports 93% on Cybench, 61.7% on DeepSWE v1.1 and 59.9% on AutomationBench. On the Dense 200 visual-grounding test it reports 42% against 41% for GPT-6-Astra. All of these figures come from Mistral's post. The post gives no before-and-after comparison against Mistral's previous flagship. A one-point margin on a single test is within the range where evaluation setup can change the ranking, so the figures should be read as claims until outside evaluators reproduce them.

Pricing is the number you can model

Mistral lists API pricing at $1.36 per million input tokens and $4.18 per million output tokens. That is the only cost figure available to buyers today. Self-hosting changes the arithmetic, because a model with 49 billion active parameters still needs hardware to hold a trillion-parameter checkpoint, and neither source states that requirement. Cost per completed task, not cost per token, will decide whether hosting the weights beats paying the API.

The capital behind the launch

Tech.eu reports that Mistral closed a Series D in September of roughly $3.4 billion at a valuation of about $24 billion. Those are the outlet's approximate figures, and the sources reviewed give no round terms beyond them. Large 4 is the first flagship release to follow that round, so it is an early public test of what the money has produced. Tech.eu also notes that Mistral has started offering third-party models to customers.

What has to happen before anyone commits

Three things settle the question. First, whether the weights ship by the end of October, and under what license, since neither source states one. Second, whether third-party evaluations match Mistral's cybersecurity and coding scores once outsiders can run the model. Third, whether measured latency and cost per task justify self-hosting over the preview API. Until the first is answered, 'open-weight' describes an intention rather than a product.

Verdict: treat Mistral Large 4 as a candidate to test, not as a standard to adopt. The confirmed facts are a preview, a published price and a dated promise of weights. Until the weights and license appear and outside benchmarks agree with Mistral's, procurement and architecture decisions should wait, and any pilot should be priced on the API rate rather than on the launch claims.