Mistral AI

Mistral Large 4

Mistral published open weights for improved Mistral Large 4 (Le Chonk) on 2026-10-29 — 1.05T MoE, 49B active, Mistral Ecosystem Open-Weight license, $0.70/$2.10 (cached $0.07). Independent HLE 35.03 and TB4 22.73 are preview-checkpoint readings; Mistral's ~41% TB4 claim is vendor-only.

Mistral Large 4 benchmarks and pricing, every number sourced: 3 tracked Mistral Large 4 benchmark scores (2 independently run, 1 still resting on a vendor’s own claim), priced at $0.70 per million input tokens and $2.10 per million output.

Mistral Large 4 architecture: MoE, natively multimodal (text + image input); 1.05T total parameters (49B activated per token); 1M-token context window.

Released
2026-10-06
Availability
Open weights published 2026-10-29 under the Mistral Ecosystem Open-Weight license; API live since the 2026-10-06 public preview, now $0.70/$2.10 per 1M tokens (cached $0.07).
License
open-weights
Context window
1M tokensvendor-stated (docs.mistral.ai model page: "Context 1M"); Artificial Analysis's comparison summary separately lists 524k — the vendor spec is recorded here, AA's figure noted on the page
Knowledge cutoff
Not disclosed
Verified
sources checked 2026-10-06–2026-10-15
Parameters
1.05T (49B active)
Architecture
MoE, natively multimodal (text + image input)

Mistral Large 4’s verified record

Mistral Large 4’s most-compared rival is Gemini 4 Argon: 2 trails and 1 not callable across their 3 shared comparisons. Mistral Large 4 is announced at $0.70/$2.10 per 1M tokens (in/out) vs Gemini 4 Argon’s $4.00/$20.00.

Against the 85 head-to-head comparisons Mistral Large 4 shares with other tracked models:

41 real gaps6 inside the band38 not called

A gap counts for Mistral Large 4 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Mistral Large 4 trails on 39 of them.

HLE · no tools Reasoning

±2 is noise
Behind
Qwen3.8-Flash-Next −3.0 · GPT-6 Luna −3.5 · MiniMax M3 −4.0 · GLM-5.3-Flash −4.9 · DeepSeek V4 Pro (0813) −6.0 · GLM-5.3 −7.3 · Grok 4.7 −8.1 · DeepSeek V4.1 Flash −8.3 · Muse Spark Open −8.9 · Claude Haiku 5.5 −9.4 · Step 5 Preview −11.4 · Kimi K3 −11.9 · Gemini 3.1 Pro Preview −12.0 · Gemini 3.8 Flash −12.8 · MiniMax M3.1 Flash −12.9 · Muse Spark 1.3 −13.7 · MiMo-V2.6-Pro −14.3 · Muse Wonder −15.1 · Kimi K3.1 Pro −17.4 · Claude Sonnet 5.5 −20.0 · Gemini 4 Argon −22.1 · Claude Opus 5.5 −26.4 · 16 previous: DeepSeek V4 Flash (0731) −3.6 · GPT-5.6 Luna −4.5 · GLM-5.2 −6.1 · Claude Sonnet 5 −6.3 · Grok 4.6 −7.9 · Qwen3.8-Max −8.1 · Muse Spark 1.2 −10.4 · Gemini 3.7 Flash −12.9 · GPT-6 Sol −12.9 · Claude Opus 4.8 −13.7 · GPT-5.6 Sol −14.5 · GPT-6.1 Sol −17.9 · GPT-6 Astra −19.6 · Claude Opus 5 −19.9 · Claude Fable 5 −20.5 · Claude Fable 5.1 −24.1
Tie
1 model within ±2
Unverified
1 model — vendor-reported on one side

Terminal-Bench 4.0 · high Terminal ops

±12.4 is noise
Behind
Tie
5 models within ±12.4

No verdict for Mistral Large 4 anywhere on DeepSWE ( nothing independently confirmed on both sides).

  • None of Mistral Large 4’s coding comparisons are independently confirmed on both sides yet.

Mistral Large 4 announced pricing

$0.70 in / $2.10 out per 1M tokens — official pricing source

What Mistral Large 4 costs per job

Mistral Large 4 cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$0.091
A codebase review1,000K / 100K$0.910
A day of agent work10,000K / 1,000K$9.10

Computed from Mistral Large 4’s list rates above — cache discounts and batch tiers are not applied.

Mistral Large 4 benchmark scores

Mistral Large 4 benchmark scores, provenance, and source links
BenchmarkScore
DeepSWE[1]
Long-horizon coding · ±9.5 is noise
Terminal-Bench 4.0(high)[2]
Terminal ops · ±12.4 is noise
HLE(no tools)[3]
Reasoning · ±2 is noise

Who ran these numbers: 2 of 3 independent — vals.ai (1), artificialanalysis.ai (1); vendor self-reported (1).

  1. DeepSWE: Mistral's launch-chart figure for DeepSWE v1.1, quoted in the 2026-10-06 announcement. No independent DeepSWE run exists for this model yet; per this site's rules a self-reported row never drives a verdict.
  2. Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 22.727 at effort high. Artificial Analysis same-conditions run (no tier suffix): 26.77. Mistral's launch prose claims 28.3 on 'Terminal-Bench 4' with no tier named — not stored as a row; this cell holds the independent run.
  3. HLE: AA's own run, board row 'Mistral Large 4 Preview' (no effort tier on the row): hle 0.35032437442076 = 35.03, text-only, no tools. Read from the Mistral Large 4 Preview page on 2026-10-15.

Notes on the record

Open weights for the improved Mistral Large 4 (Le Chonk) were published on 2026-10-29 (mistral.ai weights-day blog; Hugging Face mistralai/Mistral-Large-4-1T-A52B). The license is the Mistral Ecosystem Open-Weight license — a modified MIT license; model-as-a-service use requires a separate commercial agreement. The 2026-10-06 public-preview entry remains the launch snapshot and is unchanged.

Pricing moved with the weights release: the docs pricing page now lists $0.70 input / $0.07 cached / $2.10 output per 1M tokens, up from $0.68/$2.09 at preview. Artificial Analysis's benchmark runs bill a separate priority tier; the preview-era figure ($1.36/$4.18) was not re-checked for this update.

Specs: 1.05T total parameters; 49B active per token and 52B including embeddings/output. The preview-era 49B-vs-52B discrepancy is resolved by that split, not a contradiction. Context remains vendor-stated at 1M tokens; Artificial Analysis separately lists 524k — both published, the vendor's recorded above. No knowledge cutoff is disclosed.

The launch chart remains self-reported and is unchanged. It claims DeepSWE v1.1 at 61.7%; a Coding Agent Index of 49.8% ahead of DeepSeek V4 Pro 0813 and Qwen3.8-Max; AutomationBench at 59.9%; 82% on an Artificial Analysis Cyber Index vulnerability-reproduction test (caveated because Claude Opus 5.5 and GPT-6 Astra refuse the task); and a Surge AI blind human eval at 3.74/5 behind Claude Opus 5 at 4.22.

Tracked independent rows are preview-checkpoint readings, not weights-release measurements: Humanity's Last Exam 35.03 (Artificial Analysis, no effort tier, checked 2026-10-15) and Terminal-Bench 4.0 22.73 (vals.ai, high effort). Mistral's ~41% Terminal-Bench 4.0 claim for the weights release is vendor-reported and is not recorded as an independent score. Intelligence Index was 38, #64 of 225 (checked 2026-10-06, pre-v5; not comparable to Index v5 shipped 2026-10-28; this site does not publish that composite).

Compare with

FAQ

What is Mistral Large 4?

Mistral AI's flagship: a 1.05T-parameter natively multimodal mixture-of-experts model, 49B active per token (52B including embeddings/output), with a vendor-stated 1M-token context. It entered public preview on 2026-10-06 and its open weights were published on 2026-10-29 under the Mistral Ecosystem Open-Weight license; API pricing is $0.70/$2.10 per million tokens.

How much does Mistral Large 4 cost?

The standard regional rate is $0.70 per 1M input tokens and $2.10 per 1M output tokens, with cached input at $0.07 (docs.mistral.ai pricing page), up from $0.68/$2.09 at the October 6 preview. Artificial Analysis's benchmark runs bill a separate priority tier; its current rate was not re-checked for this update.

Has Mistral Large 4 been independently benchmarked?

Partially, and only on the preview checkpoint. Two of this site's fourteen tracked benchmarks carry an independent row: Humanity's Last Exam at 35.03 (Artificial Analysis, no effort tier shown, checked 2026-10-15) and vals.ai's Terminal-Bench 4.0 run at 22.73 (high effort). Both predate the 2026-10-29 weights release. Mistral's ~41% Terminal-Bench 4.0 claim for the weights checkpoint is vendor-reported and is not recorded as an independent score. Artificial Analysis also ran the preview in the Intelligence Index at 38 (#64 of 225, checked 2026-10-06); that is a pre-v5 reading, not comparable to Index v5 shipped 2026-10-28, and this site does not publish a composite.

Is Mistral Large 4 available?

Yes. The API has been live since the 2026-10-06 public preview via Mistral Studio, and the open weights were published on 2026-10-29 at Hugging Face (mistralai/Mistral-Large-4-1T-A52B) under the Mistral Ecosystem Open-Weight license.

What context window does Mistral Large 4 have?

Mistral's docs state the context as 1M tokens. Artificial Analysis's comparison summary separately lists a 524k window for the model — both figures are published; this page records the vendor's spec and notes the difference.

Is Mistral Large 4 open source?

Its weights are now open, but the license is not plain MIT: the Mistral Ecosystem Open-Weight license is a modified MIT license, and model-as-a-service use requires a separate commercial agreement. This site's license field therefore reads open-weights.

How does Mistral Large 4 compare with other frontier models?

Mistral's launch chart is still its own claim: a Surge AI blind eval scored the preview 3.74/5, second behind Claude Opus 5 (4.22) and ahead of Kimi K3 (3.59). The chart also claims DeepSWE v1.1 61.7% and leads on coding-agent benchmarks. Two independent tracked scores exist, both from the preview checkpoint — Artificial Analysis's 35.03 on Humanity's Last Exam and vals.ai's 22.73 on Terminal-Bench 4.0. Mistral's ~41% Terminal-Bench 4.0 claim for the weights release is not recorded as an independent score.

What is Mistral Large 4's knowledge cutoff?

Not disclosed. Neither the launch announcement nor the docs model page states a training-data cutoff date, so this page leaves the field blank rather than borrowing an estimate.

What changed on 2026-10-29?

Mistral published the open weights for the improved Mistral Large 4 checkpoint under the Mistral Ecosystem Open-Weight license, and API pricing moved to $0.70/$2.10 per 1M tokens (cached $0.07). The 2026-10-06 preview entry remains unchanged; the existing independent scores were run on that preview checkpoint, not on the weights release.

Further reading

Benchmark guides

  • DeepSWE — what Mistral Large 4’s coding score on it does and doesn’t prove.
  • Terminal-Bench 4.0 — what Mistral Large 4’s agentic score on it does and doesn’t prove.