What’s new, in 20 seconds
- Mistral released Mistral Large 4 as a public preview on October 6, 2026: about 1.05 trillion total parameters, 52B active, plus a 1.6B vision encoder. Its docs list a 1M-token context and $1.36 input, $0.14 cached input and $4.18 output per million tokens.
- Mistral says it will release the weights 'by the end of the month'; its Hugging Face placeholder shows an ETA of October 31, 2026. No license has been named, and Mistral's EU technical document for the model still lists it as proprietary until weights are released.
- Artificial Analysis gave the preview 38 on its Intelligence Index on October 6, level with GPT-6 Luna (max). On Vals AI it scores 27.78% on Terminal-Bench 4.0, ranked 22 of 54 on October 23, close to Mistral's own 28.3%.
Contents
Mistral Large 4 is Mistral AI's new flagship: a mixture-of-experts (MoE) model with about 1 trillion parameters, 52 billion of them active per token, released as a paid API preview on October 6, 2026. Mistral calls it "open-weight," but the weights aren't out yet. Its Hugging Face placeholder gives October 31, 2026 as the release date, which would make it one of the largest open models available to download. Until then it is a hosted model, and most published results are Mistral's own.
This story is based on Mistral's announcement, model docs, changelog and EU technical documentation, its Hugging Face placeholder page, and results published by Artificial Analysis, Vals AI and Arena, checked October 23, 2026. Our DeepSeek V4.1 Flash story first noted the preview on October 9.
The short version
- Size: about 1.05 trillion total parameters, 52B active (49B excluding embedding and output layers), plus a 1.6B vision encoder. Text and image in, text out, per Mistral's EU technical document.
- Price: $1.36 input, $0.14 cached input and $4.18 output per million tokens. A 50% launch discount ran for the first two weeks.
- Context: Mistral's docs say 1M tokens; Artificial Analysis and Vals list about 512K. Mistral hasn't explained the gap.
- Weights: promised "by the end of the month"; Hugging Face shows an ETA of October 31. No license yet.
- Independent results: mid-pack. Artificial Analysis scores it 38, level with GPT-6 Luna (max); Vals ranks it in the middle or lower half on most of its tests.
What is Mistral Large 4?
Mistral's announcement, nicknamed "Le Chonk," describes a "natively multimodal" hybrid instruct-and-reasoning MoE trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European datacenters. The docs (API name mistral-large-4) list function calling, structured outputs, batching and built-in tools.
How much does Mistral Large 4 cost?
| Model | Input ($/M tokens) | Cached input | Output ($/M tokens) | Weights |
|---|---|---|---|---|
| Mistral Large 4 (preview) | $1.36 | $0.14 | $4.18 | Promised for October |
| Claude Opus 4.8 | $5.00 | $25.00 | No | |
| Claude Opus 5.5 | $4.00 | $20.00 | No | |
| Claude Sonnet 5.5 | $2.00 | $10.00 | No | |
| DeepSeek V4 Pro (peak) | $1.32 | $0.044 | $3.96 | |
| DeepSeek V4.1 Flash (peak) | $0.30 | $0.006 | $1.20 | Yes (MIT) |
Table 1Mistral Large 4 lists about a sixth of Claude Opus 4.8's output price, but almost the same as DeepSeek V4 Pro at peak hours. DeepSeek's off-peak rates are half the peak rates. Source: Mistral docs, Anthropic pricing docs, DeepSeek API docs.
Last verified: October 23, 2026
Mistral's changelog says launch pricing was 50% off for two weeks from October 6. Artificial Analysis put the discounted rate at $0.68 input and $2.09 output. The docs now show the full list price. Per-token prices don't tell the whole story: Artificial Analysis estimated $1.13 per task to run its Intelligence Index, against $0.27 for DeepSeek V4.1 Flash (max). For cheaper options, see our comparison of cheap, fast models.
Is Mistral Large 4 better than Claude Opus 4.8?
No published evaluation compares the two directly. Mistral's own blind coding evaluation, run by Surge AI on a 1–5 scale, scored the preview 3.74, behind Claude Opus 5 at 4.22 and ahead of GLM-5.3 (3.60) and Kimi K3 (3.59). For how the current US frontier models compare, see our Gemini 4 Argon, Claude Opus 5.5 and GPT-6 comparison.
What benchmark results does Mistral report?
All of these are vendor-reported by Mistral on October 6:
- DeepSWE v1.1: 61.7%; SWE-Atlas-QnA: 59.4%; Terminal-Bench 4: 28.3%
- AutomationBench (657 business workflows): 59.9%
- Cybench (40 exercises): 93%; a reproduce-and-patch test: 82%
- Lakera's B3 agent security benchmark: resists 93.3% of attacks
Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on its reproduce-and-patch test because they refuse, which reflects safety policy as much as ability.
What do independent benchmarks show?
- Artificial Analysis (October 6, 2026): Intelligence Index 38 (version 4.3.2), "comparable to GPT-6 Luna (max, 38) and DeepSeek V4.1 Flash (max, 39)." It called Large 4 "the most intelligent model from outside the US and China." On October 23 its model page listed 87.0 output tokens per second and a 2.77-second time to first token on Mistral's API; those measurements change over time. DeepSeek has since released V4.1 Flash 1017, which Artificial Analysis scored 45 on October 20.
- Vals AI (model page, checked October 23): run at high reasoning effort. Terminal-Bench 4.0: 27.78% (22nd of 54). Vibe Code Bench v1.1: 78.40% (29th of 118). Finance Agent v2.1: 61.38% (15th of 30). SWE-bench Multimodal: 33.13% (20th of 20). Ranks shift as Vals adds models. The page doesn't date individual runs.
- Arena: its changelog shows Large 4 added to Code Arena: WebDev on October 7 and to the Text and Agent leaderboards on October 8. Arena's leaderboard page didn't show its rank as of October 22, and the changelog entries give no scores.
Vals' Terminal-Bench 4.0 result (27.78%) is close to Mistral's own 28.3%. Each evaluator uses its own harness and settings, so their results can't be merged into one ranking.
When will Mistral Large 4's open weights be released?
Mistral Large 4, from preview to promised weights
- Oct. 6, 2026Public preview on Mistral's API at $1.36/$4.18, 50% off for two weeksMistralConfirmed
- Oct. 7–8, 2026Added to Arena's WebDev, Text and Agent leaderboardsArena changelogConfirmed
- Oct. 31, 2026Open weights on Hugging Face (placeholder ETA)Hugging FaceExpected
- Confirmed
- Expected
Mistral's announcement says: "We will release the weights by the end of the month." The Hugging Face repository mistralai/Mistral-Large-4.0-1T05-A52B is an "Upcoming release" page with an October 31, 2026 ETA. It lists no files, license or context length.
The license matters. Mistral has used Apache 2.0 for some models and a revenue-capped "Modified MIT" license for others, including Mistral Medium 3.5 and its Timestral forecasting models. Our open-weight licenses guide explains the difference. Mistral's EU technical document lists the model as proprietary and says it "will be updated in case of open weights release."
Where can you use Mistral Large 4?
On Mistral's API through Mistral Studio, which runs the preview on European infrastructure. The technical document also lists Le Chat and partner-served deployments as channels. Mistral says the model will be "available across multiple regions worldwide" but hasn't named cloud partners. Self-hosting a model of this size needs a multi-GPU server; our cost guide explains when that beats paying per token.
What's confirmed, reported and uncertain
- Confirmed (Mistral): October 6 preview, about 1.05T total and 52B active parameters, $1.36/$0.14/$4.18 pricing, two-week launch discount, the "end of the month" weights promise.
- Vendor-reported: all of Mistral's benchmark scores and its Surge AI evaluation.
- Independent: Artificial Analysis' index of 38 and cost per task; Vals' results.
- Uncertain: the license; whether the weights land on October 31; the context window (1M per Mistral, about 512K per the evaluators); cloud availability.
About this story
Based on the sources linked below. How we report: Editorial standards.
Comments (0)