Mistral Large 4 review, pricing and verdict

Mistral Large 4 launched on 6 October 2026 and scores 38 on the Artificial Analysis Intelligence Index while generating 116.1 output tokens per second. It suits teams that want an open-weight model for security, document and agent work, but preview status and an unannounced license mean production self-hosting has to wait.

Mistral Large 4 is Mistral AI's 1.05 trillion parameter multimodal Mixture-of-Experts model, released on 6 October 2026 as a public preview. Mistral reports 93% on Cybench and 61.7% on DeepSWE v1.1. It activates 52 billion parameters per token, accepts text and images, and its weights are promised for the end of October 2026.

Provider: Mistral AI · Family: Mistral Large

More about Mistral AI on HokAI

Context window: 524,288 tokens · Max output: 262,144

Input modalities: text, image · Output: text, tool-calls

About Mistral Large 4

Mistral Large 4 is the newest model in Mistral AI's Large line, launched on 6 October 2026 as a public preview and nicknamed "le Chonk" by the company. It is a granular Mixture-of-Experts model with 1.05 trillion total parameters, 52 billion active per token and a 1.6 billion parameter vision encoder, according to Mistral's model documentation. Mistral calls it its largest and most capable model to date and says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. It follows Mistral Large 3, the 675B total, 41B active model released in December 2025, and is built by Mistral AI.

Mistral publishes most results as charts, with the key figures in the announcement text, so every number in this paragraph is vendor-reported unless it names another source. On agentic coding Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, for a combined Coding Agent Index of 49.8% that it places ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On AutomationBench, which covers 657 business workflows across apps such as Gmail, Slack and Salesforce, it reports 59.9%, and it reports 1,393 Elo on AA-Briefcase. In cybersecurity it reports 93% on Cybench, a set of 40 security competition challenges, and 82% on CyberGym-E2E, a test that asks a model to reproduce a real vulnerability and then patch it. A blind Surge AI human evaluation of coding output rated it 3.74 out of 5, second of five models behind Claude Opus 5 at 4.22 and ahead of Kimi K3 at 3.59 and GLM-5.3 at 3.60. Artificial Analysis, which tested the preview independently, scores it 38 on its Intelligence Index v4.3.2 and measures 116.1 output tokens per second, but notes that it is very verbose, generating 200 million output tokens to complete the index.

Mistral's documentation card lists a 1M-token context window, but both Artificial Analysis and OpenRouter list the live preview endpoint at 524,288 tokens with a 262,144-token output cap, so plan around the lower figure until Mistral clarifies. Input is text and images and output is text. The API supports function calling, structured outputs, document question answering, prefix completion, batching, and the Agents and Conversations endpoints with built-in tools. Reasoning can be switched: OpenRouter lists reasoning effort values of high and none, with high as the default. Mistral says a significant share of the training data was multilingual across more than 160 languages, including every official EU language.

At launch the model is served through Mistral's own API in Mistral Studio, and OpenRouter lists Mistral as its only provider. Mistral has promised to release the weights by the end of October 2026 with more detail on the architecture, and no license has been announced, so self-hosting is not possible today. Only about 5% of the parameters activate per token, which sets compute cost, but the full 1.05 trillion parameters still have to sit in memory once weights ship. Mistral also describes a European deployment that it operates end to end under European law.

On safety, Mistral reports that the model resists 93.3% of attacks on Lakera's B3 agent security benchmark and scores 1.691 of a possible 2 on the KORA benchmark, and says its refusal rate on harmful cyber prompts from JailbreakBench, StrongREJECT and AgentHarm is higher than any open-weight model it compared. It adds that Claude Opus 5.5 and GPT-6 Astra score near zero on one cyber test because they refuse the task, which is Mistral's claim and not an independent finding. Before weights ship, Mistral is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who access it with reduced moderation. No system card was linked from the announcement.

Mistral Large 4 fits teams that want open-weight control for security research, document and image analysis, and agent workflows, and Mistral highlights finance and legal results measured by Vals.ai, where it says the model exceeds GPT-6 Astra. It is the wrong choice today for anyone who needs downloadable weights this month, a generally available endpoint with a stable version, or tight output-token budgets. Teams that need open weights now can use Mistral Large 3, which Mistral lists under Apache 2.0.

Pricing

The preview API is billed at a sale price of $0.68 per 1M input tokens, $0.07 per 1M cached input tokens and $2.09 per 1M output tokens. Mistral's pricing page shows these against crossed-out list prices of $1.36, $0.14 and $4.18, which is exactly 50% off, and states no end date for the sale. Artificial Analysis lists the list prices with a 90% cache discount. For comparison, Mistral Large 3 is listed at $0.50 input and $1.50 output per 1M tokens. Batch and Priority rates sit on separate tabs of Mistral's pricing page and are not recorded here.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Granular Mixture-of-Experts: Sparse expert routing keeps compute per token far below the total parameter count, and a separate vision encoder handles native image input.
• Long served context: The preview endpoint serves a window below the 1M shown on Mistral's docs card, with a separate output cap for long completions.
• Switchable reasoning: A hybrid instruct-and-reasoning design with two effort settings, high (the default) and none, as listed by OpenRouter.
• Agent and tool support: Function calling, structured outputs, document question answering, batching, and Agents and Conversations endpoints with built-in tools.
• 160+ language training data: Mistral says a significant share of training data was multilingual, covering every official language of the European Union.

Pros

• Strong vendor-reported cyber results, with weights promised so security teams can self-deploy.
• Supports image input with function calling and structured outputs in a single model.
• Mistral reports 93.3% resistance on Lakera's B3 agent security benchmark.

Cons

• Weights, license and architecture details are unpublished at launch, so self-hosting is not yet possible.
• Artificial Analysis calls the model very verbose, so per-task cost can run high.
• Preview status: Mistral says the reinforcement learning run behind it is still in flight and the model may change.

Benchmarks

• AA Intelligence Index: 38 cited: Artificial Analysis · 08 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• Output speed: 116 tok/s cited: Artificial Analysis · 08 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Mistral Large 4 Official Page