Benchmarks
More
OverviewBenchmaxxingCompareChartsCost vs CapabilityProviders per ModelProvider explorerGatewaysEU & SovereignJev-class modelsImage JevBenchAudioJevBenchSubmit a modelAbout
← All models

Devstral Small 2

open weightsdeprecated by benchmark source
Mistral · released 2025-12-09 · 0 offers

Context 256K tokens

Providers

No provider publishes an API price for this model yet, so no cost can be modeled.

Composite

2 of 7 inputs6 radar axes: DesignArena's two boards share one

13.1

 Thin data · 2/7Based on only 2 of 7 Composite inputs — treat this rank as uncertain
AA Coding v4.xCoding Agent v1.4AA Intelligence (est.)AA AgenticEpoch ECISoftware ECIDesignArenaAA Coding v4.x: percentile 35AA Intelligence (est.): percentile 27AA Agentic: percentile 34

AA Coding v4.x 29.3Coding Agent v1.4 —AA Intelligence (est.) 3.3AA Agentic 4.8Epoch ECI —Software ECI —DesignArena —/—

Radar: percentile among all models measured on each input; a gap means not measured.

Benchmark sheet

23 of 319 registered benchmark versions · bars show the percentile among all models measured on each benchmark.

1 of 25 values are Mistral's own claims (†), not independent measurements; a matching independent result replaces a claim as soon as one exists.

Compare this model →

Agentic

  • AA-Briefcase23496 Elo

    AA-Briefcase published 2026-09-10 → · Published board — Tests multi-week professional knowledge-work projects with linked tasks and large source collections.

  • GDPval-AA v232680 Elo

    GDPval-AA v2 v2 → · Published board — Tests professional knowledge-work deliverables across occupations using AA's Stirrup harness.

  • Terminal-Bench Hard (AA)5416.7%

    Terminal-Bench Hard (AA) pinned revision 74221fb → · Published board — Tests a pinned 44-task hard subset of terminal-based work using Terminus 2.

  • Terminal-Bench v2.1 (AA)3529.6%

    Terminal-Bench v2.1 (AA) v2.1 → · Published board — Tests terminal-based work on the 89-task verified refresh using Terminus 2.

  • Terminal-Bench v4.0 (AA)180.0%

    Terminal-Bench v4.0 (AA) v4.0 → · Published board — Tests terminal-based work on the 66-task release using mini-SWE-agent v2.4.6.

  • AA Agentic Index344.80

    AA Agentic Index published 2026-09-18 → · Published board — Artificial Analysis publishes one Agentic Index per model, measured on its primary configuration and relayed by OpenRouter's Benchmarks API; the value is attached at family scope on the deterministic representative. Artificial Analysis publishes this board without a version number; we keep the date each result was retained.

Coding

  • SciCode (AA subproblems) v1.0.11132.4%

    SciCode (AA subproblems) v1.0.1 → · Published board — Tests scientific Python programming with scientist-annotated background information.

  • SWE-bench Verifieddeveloper's claim68.0%† self-reported by the developer

    SWE-bench Verified published 2026-09-10 → · vendor protocol: overall; Self-reported vendor headline figure; evaluation harness not specified in source — A 500-instance human-filtered subset of SWE-bench created with OpenAI, served as the default Verified leaderboard view.

  • AA Coding Index v4.x · frozen 2026-10-273529.3

    AA Coding Index v4.x · frozen 2026-10-27 → · Published board — Artificial Analysis withdrew the Coding Index with Intelligence Index v5.0 (2026-10-28). These are the last v4.x values, frozen with their own date (2026-10-27).

Instruction-following

Knowledge

Long-context

  • GDP.pdf (AA)91.0%

    GDP.pdf (AA) published 2026-09-10 → · Published board — Tests professional reasoning over long PDFs with AA document preparation and grading.

  • AA-LCR v1.13028.0%

    AA-LCR v1.1 v1.1 → · Published board — Tests reasoning across multiple long documents with corrected answer keys and grading.

Math

  • AIME 2025 (AA)3434.3%

    AIME 2025 (AA) v2025 → · Published board — Advanced mathematical problem solving on AIME I and II 2025.

Science

Tool-use

  • AutomationBench-AA v1.0.6253.1%

    AutomationBench-AA v1.0.6 → · Published board — Tests multi-app SaaS workflows through REST tools on a held-out AutomationBench split.

  • τ²-Bench Telecom (AA)2623.4%

    τ²-Bench Telecom (AA) published 2026-09-10 → · Published board — Tests dual-control telecom agents that coordinate tool use with a simulated user.

  • τ³-Banking (AA) v1.0.13210.7%

    τ³-Banking (AA) v1.0.1 → · Published board — Tests banking support agents that retrieve policies and change account state through tools.

Uncensored

  • UGI Willingness966.80

    UGI Willingness published 2026-09-10 → · Published board — A private-question evaluation of how far a model follows challenging instructions before refusing or deviating.

  • UGI7245.3

    UGI published 2026-09-10 → · Published board — A private-question suite combining sensitive-topic knowledge with willingness to follow controversial instructions.

Vision

Writing

  • UGI Writing3931.6

    UGI Writing published 2026-09-10 → · Published board — A maintainer-owned writing evaluation considering intelligence, style, repetition and length adherence, informed by human preferences.

Unusual results

1 threshold-crossing signal flagged · 23 eligible benchmark families

How flags are calculated

Heuristic screen, not statistical significance: benchmark families are correlated and source uncertainty is unknown. Peer evidence needs ≥ 20 independently measured matched configurations from ≥ 10 distinct model families. A flag needs a directed population z-score of magnitude ≥ 1.5 and a gap of ≥ 1.5 from the leave-one-benchmark-family-out mean z in the same direction, over ≥ 5 other benchmark families.

Missing coverage · 296 benchmark versions

No result does not mean a zero, or that the model was never tested. Collection failures and disputed versions retain their distinct status.

Devstral Small 2 — benchmarks & cost | Benchmark Heaven