Mistral Large 4 Benchmarks: Official and Independent Results

The short version

Mistral Large 4 is strongest at security work, legal and finance agents, and pointing at objects in images; it is average on broad intelligence indexes. That pattern holds across both Mistral’s own charts and the independent leaderboards.

• Security: 82% on CyberGym-E2E and 93% on Cybench, ahead of every open-weight model Mistral compared against.
• Legal: 15.8% on Harvey’s Legal Agent Benchmark, ranked #6 of 75 models by Vals AI, above GPT-6 Astra (5.4%).
• Coding agents: 61.7% on DeepSWE v1.1, above DeepSeek V4 Pro 0813 (57) and Qwen3.8 Max (51), below Kimi K3 (68).
• Vision: 42.0 on Dense200 grounding, slightly above GPT-6 Astra (41.5).
• Overall: Artificial Analysis gives it 38, behind Claude Opus 5.5 (58) and GPT-6 Astra (53) but ahead of DeepSeek V4 Pro (36.0).

Coding and agent benchmarks (Mistral-reported)

On coding agents Mistral Large 4 beats DeepSeek V4 Pro and Qwen3.8 Max but trails GLM-5.3 and Kimi K3 on some tests. Mistral ran these with Artificial Analysis harnesses; each rival used its own preferred agent tool (for example Claude Code for Qwen3.8 Max and Codex for DeepSeek), so treat small gaps as close calls.

Mistral reports a combined Coding Agent Index of 49.8%. In the Surge AI blind human evaluation of coding answers, raters scored it 3.74 out of 5, ahead of Kimi K3 (3.59) and GLM-5.3 (3.60) but behind Claude Opus 5 (4.22).

Cybersecurity and safety benchmarks

Cybersecurity is where Mistral Large 4 stands out most. On CyberGym-E2E, which asks a model to reproduce and then patch real vulnerabilities, it scores 82% while Mistral reports that Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the tasks. On the Artificial Analysis Cyber Index it completes 50 tasks, against 33 for GPT-6 Astra and 29 for Claude Opus 5.5.

It also refuses clearly harmful requests: an average of 95.3 across JailbreakBench, StrongREJECT and AgentHarm, and 93.3% attack resistance on Lakera’s B3 agent security test. Mistral gives cyber leaders, vetted partners and state authorities a reduced-moderation build for red-teaming before the weights ship.

Finance, legal, science and vision

For professional work, Mistral Large 4 leads clearly on legal agents and is competitive on finance and science. Its 15.8 on Harvey’s Legal Agent Benchmark is about three times GPT-6 Astra’s 5.4. On Finance Agent v2 it scores 54.7, just above GPT-6 Astra (53.5) and just below GLM-5.3 (55.8).

On images it reads charts and locates objects well. Dense200 measures how precisely a model draws bounding boxes, and Mistral Large 4 scores 42.0 against 41.5 for GPT-6 Astra and 28.9 for Kimi K3. On ChartQA Pro it scores 63.1, slightly under GPT-6 Astra (65.1).

Independent results: Artificial Analysis

Artificial Analysis scores Mistral Large 4 at 38 on its Intelligence Index v4.3.2, placing it #64 of 225 models and 8th among open-weight models. It ranks ahead of DeepSeek V4 Pro (36.0) and GLM-5.2 (33.7), and behind GLM-5.3 (45) and MiMo-V2.6-Pro (46).

It is fast but talkative. Output speed is 116.1 tokens per second against a median of 87, yet it used 200M output tokens to finish the index against a median of 81M, so long answers add to cost and wait time. Artificial Analysis currently lists it as a proprietary model because the weights are not public yet.

Independent results: Vals AI

Vals AI places Mistral Large 4 at 48.05% on the Vals Index v2.1, #32 of 44 models, almost level with Qwen 3.8 Max (48.27%) and DeepSeek V4 Pro 0813 (47.63%). Gemini 4 Argon leads that index at 68.90%, followed by Claude Opus 5.5 at 66.97%.

Its best Vals placements are legal and finance work; its weakest are formal proofs and puzzle-style science reasoning.

How to read these numbers

Vendor scores and independent scores answer different questions, and the gaps between them are worth knowing before you choose a model.

• Terminal-Bench 4.0 differs by harness: 28.3% in Mistral’s Artificial Analysis run versus 22.73% on Vals AI.
• Mistral calls its SciCode-Verified result state of the art among open-weight models, yet its own chart shows MiMo-V2.6-Pro (91.9), GLM-5.3 (92.5) and Qwen3.8 2.4T A95B (93.8) above its 91.8.
• The internal expert win rate against GLM-5.3 is 68% for STEM on the domain chart and 69% on the breakdown chart; CAD is 62%, finance 50% and code 48%.
• All Mistral numbers are for the preview build released on Oct 6, 2026; scores can change once the final weights ship.
• Reasoning effort changes cost more than quality for some users: on Hacker News, Simon Willison reported little difference between reasoning_effort "high" and "none" in his test.

Try the model on your own prompts

Benchmarks show averages; your prompts are what matter. You can send Mistral Large 4 your own coding, legal or finance questions in the chat on our homepage, free for 3 messages a day without an account. Token costs for this level of performance are broken down on /api-pricing, and the /vs-deepseek-v4 and /vs-mistral-large-3 guides go head to head.