Mistral Large 4
Mistral's new flagship, a 1.05-trillion-parameter mixture-of-experts model with 52B active parameters and a vision encoder, released as a public preview on October 6, 2026. Mistral calls it open-weight, but the weights had not been published as of this snapshot.
Mistral AILanguage modelReleased October 2026Data checked October 18, 2026
This model is not in the current people's rankings: it is either new and still collecting votes, or ranked below the top 12. See the AI leaderboard for the models people rate highest.
On paper
Fixed-test rankings from the benchmarks tab: Speed #14 (tied), Reasoning #22.
Strengths and weaknesses
- 1M-token context window and image input
- Moderate regular price ($1.36 / $4.18 per million tokens), with half-price launch rates at the time of this snapshot
- Artificial Analysis Intelligence Index of 38.4, level with GPT-6 Luna and Muse Spark Open
- Near the bottom of this list on the LMArena text and creative writing leaderboards
- Still a preview, so behavior and pricing may change before general release
Best for
- Teams that prefer a European AI provider
- General chat and document work
Every score we have
Each bar compares this model with the best result any model on this page has on the same test.
- LMArena text arena (human votes)1,431.8Arena score 1431.8 (95% CI 1424.5 to 1439.0); 7,451 votes; entry "mistral-large-4"; position 111 of 416 on the full board
- LMArena creative writing1,367.7Arena score 1367.7 (95% CI 1352.6 to 1382.8); 1,694 votes; entry "mistral-large-4"; position 159 of 414 on the full board
- LMArena WebDev (coding)1,542.1Arena score 1542.1 (95% CI 1529.2 to 1555.0); 2,536 votes; entry "Mistral Large 4"; position 43 of 144 on the full board
- LMArena Agent Arena-0.040IPS score -0.0403 (95% CI -0.0597 to -0.0208); 14,030 sessions; entry "Mistral Large 4"; position 40 of 52 on the full board
- Artificial Analysis Intelligence Index38.438.4 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 38), setting "Mistral Large 4 Preview"
- Humanity's Last Exam (Artificial Analysis)35.0%35.0% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "Mistral Large 4 Preview"
- Terminal-Bench 4.0 (Artificial Analysis run)26.8%26.8% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "Mistral Large 4 Preview"
- Long-context reasoning (AA-LCR)81.3%81.3% on AA-LCR (long-context reasoning), setting "Mistral Large 4 Preview"
- MMMU-Pro, image understanding76.4%76.4% on MMMU-Pro (image understanding, AA), setting "Mistral Large 4 Preview"
- AA-Omniscience, accuracy minus hallucination-5.3-5.3 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "Mistral Large 4 Preview"
- Output speed, tokens per second71 t/s71 output tokens per second (AA median), row "Mistral Large 4 Preview"
- Cost per Artificial Analysis test task$1.13$1.13 to run one AA Intelligence Index task, row "Mistral Large 4 Preview"
Compare with
- Claude Opus 5.5Anthropic's flagship for long-running agentic coding and knowledge work, and the model Anthropic suggests starting with for most jobs.
- Claude Sonnet 5.5Anthropic's mid-priced model, which Anthropic describes as its best mix of speed and intelligence.
- Claude Fable 5.1Anthropic's top tier for the hardest reasoning and long-horizon agent work, which Anthropic recommends when Opus 5.5 still falls short.
- Claude Haiku 5.5Anthropic's smallest and fastest current model, built for high-volume, latency-sensitive jobs like classification, extraction and routing.