Mistral Large 4: benchmarks, price, API, weights

What is Mistral Large 4

Mistral Large 4 is the largest model Mistral AI has released — « our largest and most capable model to date », in the words of its launch post of 2026-10-06. It is the successor of Mistral Large 3 and the first Large model that Mistral describes as a hybrid: the same weights answer directly or think before answering, depending on the request. Mistral nicknames it « le Chonk » (« unofficially ML4, very officially: le Chonk »).

According to Mistral, the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres, on data covering more than 160 languages, including every official language of the European Union — a point that matters for European buyers whose users do not all write in English.

Specification	Value
Architecture	Mixture-of-experts, hybrid instruct + reasoning (Mistral)
Parameters	~1T in total, 49B active per token (Mistral)
Modalities	text + image in, text out
Context window	524,288 tokens (as exposed by the API)
Maximum output	262,144 tokens (as enforced by the API)
Languages	160+, including all official EU languages (Mistral)
Function calling	yes — measured, see below
Open weights	announced for the end of October 2026 — not published yet
Licence	not published yet
Preview	2026-10-06

Two figures circulating in the press deserve a correction. The context window: some articles say one million tokens, but the API's own model list says 524,288 — that is the number your requests will hit. The licence: Mistral Large 3 shipped under Apache 2.0, but Mistral has not published the licence of Large 4's weights yet, and it would be a mistake to assume it before reading it.

Benchmarks: Mistral's claims and an independent check

The scores below are those published by Mistral in its launch post. They are a preview's numbers, chosen by the vendor, and mostly given without competitor figures — so we quote only what is quantified.

Benchmark (Mistral's post)	Mistral Large 4
DeepSWE v1.1	61.7%
SWE-Atlas-QnA	59.4%
Coding Agent Index	49.8%
Terminal-Bench 4.0	28.3%
AutomationBench	59.9%
AA-Briefcase	1,393 Elo
Dense 200 (visual grounding)	42% (GPT-6 Astra: 41%)
B3 attack resistance	93.3%

The one head-to-head with numbers is a human evaluation run by Surge AI, where Mistral Large 4 scores 3.74 — ahead of Kimi K3 (3.59), GLM-5.3 (3.60) and GLM-5.2 (3.40), behind Claude Opus 5 (4.22). Mistral also says the model leads open-weight models on the Artificial Analysis Cyber Index and on Harvey's legal agent benchmark, without giving those scores.

An independent check is already available. Vals.ai evaluated the model on launch day: 48.05% on the Vals Index (32nd of 44), 54.68% on Finance Agent v2, 15.83% on Harvey's Legal Agent Benchmark (6th of 75 — a strong result on a hard benchmark), and 22.73% on Terminal-Bench 4.0, below the 28.3% Mistral reports for the same benchmark. Harness and settings differ between the two, so neither figure is « wrong » — but the gap is a reminder that a launch-day table is the vendor's best case.

Our reading: Mistral Large 4 is not the top open model on every agentic leaderboard — on DeepSWE, for instance, DeepSeek reports higher scores for its own recent models — but it is a frontier-class model from a European lab, with a strong showing on legal and finance work and a 512K context. As we say on every model page, a leaderboard score is a starting point, not a verdict: run your own hard prompts.

What we measured on the API (2026-10-06)

We wired the model on launch day and checked the API surface with real calls rather than trusting the announcement:

• Context and output limits: the model list returns a 524,288-token context; a request with max_tokens above 262,144 is refused with an explicit error.
• Tool calling: a request with a get_weather tool returned a valid tool_call with correct JSON arguments, and the role: "tool" round trip was accepted.
• Streaming: usage (cached tokens included) arrives on the final chunk, so streamed calls are metered exactly like non-streamed ones.
• A strict request schema: Mistral's API rejects, with a 422, fields that most OpenAI SDKs send by default — user, seed, max_completion_tokens, a name on a message, the developer role. On Frontière AI the gateway translates them (for example max_completion_tokens becomes max_tokens, developer becomes system), so an existing OpenAI client works unchanged.

We also ran our two Frontière Verified suites through the production gateway on launch day. Benchmark: passed — 25 of 25 QA answers correct, 0.56 s mean time to first token, 38.7 tokens/s over a 1,269-word generation. Security audit: passed — 14 of 15 adversarial prompts answered safely, all five categories passed; the one miss was a prompt injection that made the model print an injected marker before answering. The two remaining checks, licence and provenance, wait for the published weights.

Hybrid reasoning: how to turn it on

By default, Mistral Large 4 answers directly, with no thinking phase — fast and cheap for extraction, classification and routine chat. To make it reason first, send reasoning_effort. Mistral's API accepts only two values, high and none; on Frontière AI any effort other than none or minimal turns thinking on.

Mistral returns the thinking as structured blocks inside the message content, a format OpenAI clients do not expect. The gateway folds it into the convention used by the other reasoning models of the catalog: the visible answer in content, the thinking in reasoning_content, in streamed and non-streamed responses alike. Thinking tokens are real output tokens and are billed as such — so reasoning_effort is also a cost dial.

curl https://getfrontiereai.eu/api/v1/chat/completions \
  -H "Authorization: Bearer $FRONTIERE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "mistral-large-4", "reasoning_effort": "high",
       "messages": [{"role": "user", "content": "Review this contract clause..."}]}'

Pricing and availability

Mistral Large 4 is live on our catalog under the slug mistral-large-4, called with the same OpenAI-compatible request as every other model. Client price (our margin and VAT included, prepaid balance, no subscription): €2.2848 per million input tokens, €7.0224 per million output tokens, €0.2352 per million cached input tokens.

Model	Lane	Input, per million tokens (client, VAT incl.)	Output, per million tokens (client, VAT incl.)
Mistral Large 4 (~1T, multimodal, preview)	EU sovereign · Mistral AI	€2.2848	€7.0224
Mistral Small 3.2 (24B, multimodal)	EU sovereign · OVHcloud	€0.1512	€0.4704

Our cost basis is Mistral's list price for the model (1.36 USD per million input tokens and 4.18 USD per million output tokens, before tax), marked up like the rest of the catalog. The rule of thumb: Large 4 for the hard problems — long documents, multi-step agents, code review, legal and financial analysis; Mistral Small 3.2 for the high-volume routine work, at a fraction of the price.

Sovereignty and GDPR

On Frontière AI, Mistral Large 4 is served by Mistral AI itself — a French company — on its EU region, which the gateway pins explicitly (api.eu.mistral.ai) rather than leaving the region to chance. Mistral states that the preview is served on the same European infrastructure the model was trained on, in a deployment it « operates end-to-end, independently of other digital service providers and under European law ». The route is therefore marked sovereign, and every response carries x-sovereign: true; keys restricted to sovereign providers can call it.

One nuance, which we display rather than hide: Mistral keeps API inputs and outputs for 30 days to monitor abuse, unless Zero Data Retention is enabled on the account. Frontière AI itself stores neither prompts nor completions on any route — but because of that provider-side window, the model's sovereignty score is 80/100 (« EU hosted ») rather than the 90 of OVHcloud and Scaleway routes, which document no routine retention. Our sovereignty methodology explains each factor, and the GDPR vs CLOUD Act article why the operator — not the model — decides jurisdiction.

For a European company, the practical point is timing: a frontier-class model is callable from launch day on its European publisher's own infrastructure, without waiting for a sovereign cloud to add it to its catalog. Once the weights are published, sovereign clouds will be able to serve it too.

FAQ