Mistral Credits Artificial Analysis With Large 4’s 41.4% Coding Score. Artificial Analysis Lists 24.7%.

What Mistral shipped

Mistral released the open weights of Mistral Large 4, nicknamed Le Chonk, on October 29. Mistral calls it a 1-trillion-parameter, natively multimodal model. Artificial Analysis lists 1.1 trillion total parameters with 52 billion active, and a 1M-token context window.

The license is the Mistral Ecosystem Open-Weight license, which Mistral describes as a modified MIT license. You can use, modify, fine-tune and sell on top of it. The catch: a business running it as a Model-as-a-Service needs a separate commercial license.

The API price is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. The preview went public on October 6, per Vals.ai’s model page.

Mistral’s headline claim is direct: “Mistral Large 4 ranks as the best open-weight model for coding.” The post credits the jump to the 3 weeks since the preview: “Just three weeks of additional reinforcement learning have since propelled it to the forefront of open-weight coding.”

The 41.4% chart

The Terminal-Bench claim reads: “On Terminal-Bench 4.0, Mistral Large 4 improved by 13 percentage points to 41% compared to its public preview, taking the lead among open-weight models while reducing output tokens by 30%.”

The chart under that sentence is titled “Artificial Analysis Coding Agent Index - Terminal-Bench 4.0.” It shows the preview at 28%, labeled “ML4 Preview - self reported,” and the release at 41.4%. GLM-5.3 sits just behind at 40%.

The post doesn’t say which harness, how many attempts, or which reasoning effort produced the 41.4%. Only the preview bar carries the “self reported” label.

What the independent leaderboards say

Artificial Analysis publishes its own Terminal-Bench 4.0 results. As of November 1, its evaluation page lists Mistral Large 4 (high) at 24.7% and Mistral Large 4 (medium) at 23.2%. The preview, now marked deprecated, sits at 26.8%.

Vals.ai runs every model through the same Mini-SWE-agent harness. Its leaderboard, updated October 29, has the open-weight release, listed as Mistral Large 4 OSS, at 26.26% (±5.13). The October 6 preview, listed as Mistral Large 4, scored 27.78% (±1.82).

Vals.ai’s error bar on the release is wide, ±5.13, so its 1.52-point dip is inside the noise. The fair reading is “no measurable gain,” not “it got worse.” But neither independent board comes anywhere near 41.4%. The gap between Mistral’s chart and Artificial Analysis’s page is 16.7 points.

The rivals in Mistral’s chart line up

Here is the odd part. The other bars in Mistral’s chart sit close to what Artificial Analysis and Vals.ai publish for those models. Only Mistral Large 4’s own bar is far off.

Mistral’s GLM-5.3 bar is within 2 points of both independent scores. Its DeepSeek V4.1 Flash 1017 bar is within 3. Mistral Large 4 is the only row where the launch chart and the independent boards disagree by more than 15 points.

We can’t tell from the page why. Mistral may have run Artificial Analysis’s setup itself at a different effort level, or used a number Artificial Analysis later replaced. Either way, the score on Artificial Analysis’s own page today is 24.7%, and that’s the one a buyer can check.

What the 3 weeks did buy: a cheaper run

The release isn’t a step backward. Mistral says it cut output tokens by 30%, and Vals.ai’s numbers back the direction of that claim.

Same price per token, 47% less per task, 35 minutes faster. On Vibe Code Bench v1.1, the release ranks #7 of 43 open-weight models. On the Vals Index it scores 56.42%, #17 of 30 overall and #5 of 15 open-weight models.

Mistral’s other coding evidence also holds up as stated. It leads Cognition’s FrontierCode 2.0 chart at 19.7%, just ahead of DeepSeek V4.1 Flash at 19.3%. In Surge’s blind pairwise tests, graders preferred it over GLM-5.3 and Kimi K3 64% of the time. Against Claude Opus 5.5, graders picked Large 4 22% of the time.

Where it ranks on price and score

On Vals.ai’s Terminal-Bench 4.0, Mistral Large 4 OSS is #26 of 59 models and #8 of 22 open-weight models. Cost per task puts it in a tough spot.

DeepSeek V4.1 Flash 1017 scores 5.56 points higher for about a fifth of the cost per task. GPT-6.1 Sol, a closed model, costs less per task than Large 4 and scores more than twice as high.

What this means if you code with AI

Read the label on every bar. A chart titled with a third party’s name isn’t the third party’s page. Before you pick a model on a launch-day number, open the leaderboard it credits and find the row yourself.

For terminal agents, Large 4 isn’t the open-weight pick. If your agent runs shell commands, edits files and fixes failing tests, GLM-5.3 and DeepSeek V4.1 Flash 1017 beat it on both independent boards, and DeepSeek does it at $0.54 a task.

It’s a real option for app building and self-hosting. 81.56% on Vibe Code Bench v1.1 is a strong open-weight result, and open weights under a modified MIT license matter if you need to self-host. Check the Model-as-a-Service clause if you plan to resell access.

Don’t pick one model for everything. The cheapest model that passes your task changes from one benchmark to the next. A setup that lets you switch models per task beats betting on any single launch post.

FAQ

Where MegaBrain fits

Full disclosure: we sell an AI coding agent, and we did not run these benchmarks. MegaBrain’s terminal coding agent puts many models behind one plan: free models on Free, included models such as GLM, Kimi, MiniMax, DeepSeek, Qwen and Grok on the paid plans, Claude, GPT and Gemini from Pro, plus any model through your own API key and local models, and Auto picks one for each task. See the plans. MegaBrain is not affiliated with Mistral AI, Artificial Analysis or Vals.ai.

Sources

All pages opened November 1, 2026.

• Mistral AI — Mistral Large 4 (Le Chonk) is now open weight (announcement, charts and pricing)
• Artificial Analysis — Terminal-Bench 4.0 evaluation page
• Artificial Analysis — Mistral Large 4 (high) model page
• Artificial Analysis — Mistral Large 4 Preview model page
• Vals.ai — Terminal-Bench 4.0 leaderboard (updated October 29, 2026)
• Vals.ai — Mistral Large 4 OSS model page (released October 29, 2026)
• Vals.ai — Mistral Large 4 (preview) model page (released October 6, 2026)