Skip to content
Model Index
  • Overview
  • Models
  • Compare
  • Benchmarks
  • Finder
LabsMethodology

Benchmarks

23 benchmarks, all run by third parties rather than by the labs themselves. Each page explains what the test measures, how much weight to put on it, and where every tracked model lands.

Independent tests

A fixed set of tasks with right and wrong answers, run the same way for every model.

Human preference

Blind votes on which answer people like more. Good for feel, silent on correctness.

Saturating

The leaders are near the ceiling, so the test no longer separates the best models.

OverallCodingAgents & knowledge workReasoning & scienceKnowledge & reliabilityVision

Overall

Composite indexes and broad human preference — the quickest read on general capability.

  • Independent test

    Artificial Analysis Intelligence Index

    A weighted composite of ten evaluations (v4.3.2): agentic knowledge work 30%, coding 20%, scientific reasoning 20% and general knowledge, reliability and long-context work 30%.

    1. 1Claude Opus 5.557.6
    2. 2Claude Sonnet 5.556.0
    3. 3Claude Fable 5.153.4

    38 of 41 models · Artificial Analysis · Oct 8

  • Independent test

    Vals Index

    Agentic performance across finance, coding, legal and tax tasks (v2.1), with each sector weighted by its share of U.S. GDP.

    1. 1Gemini 4 Argon68.9%
    2. 2Claude Sonnet 5.567.0%
    3. 3Claude Opus 5.567.0%

    26 of 41 models · Vals AI · Oct 7

  • Human preference

    Arena · Text

    Which answer people prefer when two anonymous models respond to the same prompt, aggregated over millions of blind votes into an Elo-style rating (style control on).

    1. 1Gemini 4 Argon1525
    2. 2Claude Opus 5.51504
    3. 3Claude Fable 5.11501

    33 of 41 models · Arena · Oct 2

Coding

Writing, running and shipping software: terminal agents, scientific code, app building and migrations.

  • Independent test

    Terminal-Bench 4.0

    End-to-end software and systems tasks solved by an agent inside a real terminal, graded by a test suite (pass@1 over three repeats).

    1. 1Claude Sonnet 5.563.6%
    2. 2Claude Opus 5.559.6%
    3. 3GPT-6 Astra59.1%

    38 of 41 models · Artificial Analysis · Oct 8

  • Independent test

    SciCode

    Research-grade scientific programming: 288 sub-problems drawn from real physics, chemistry, biology and maths workflows, each of which must pass all unit tests.

    1. 1Claude Opus 5.566.9%
    2. 2Claude Fable 5.163.1%
    3. 3Gemini 4 Argon61.8%

    38 of 41 models · Artificial Analysis · Oct 8

  • Human preference

    Arena · WebDev

    Blind head-to-head votes on working web apps that two models build from the same prompt.

    1. 1Claude Opus 5.51814
    2. 2GPT-6 Astra1788
    3. 3Claude Sonnet 5.51773

    34 of 41 models · Arena · Oct 7

  • Independent test

    Vibe Code Bench

    Building a complete, working application end to end from a product specification.

    1. 1Claude Sonnet 5.592.4%
    2. 2Gemini 4 Argon91.9%
    3. 3Claude Haiku 5.590.4%

    26 of 41 models · Vals AI · Oct 7

  • Independent test

    Code Migration

    Porting existing projects to another programming language, including COBOL modernisation.

    1. 1Claude Sonnet 5.570.1%
    2. 2Gemini 4 Argon69.5%
    3. 3Claude Opus 5.568.9%

    26 of 41 models · Vals AI · Oct 7

Agents & knowledge work

Multi-step professional tasks with tools and files — finance, legal, tax and office deliverables.

  • Independent test

    GDPval-AA

    Real deliverables — documents, spreadsheets, presentations — across 44 occupations from OpenAI's GDPval set, judged pairwise by a panel and converted to a normalised Elo scale (v2.1).

    1. 1Claude Opus 5.568.3
    2. 2Claude Sonnet 5.567.0
    3. 3Claude Fable 5.162.9

    39 of 41 models · Artificial Analysis · Oct 8

  • Human preference

    Arena · Agent

    How well models orchestrate tools in real agent sessions, scored from behavioural signals such as task completion, user praise or complaints, steering burden, recovery steps and tool hallucinations.

    1. 1Claude Fable 5.1+14.3
    2. 2Claude Opus 5.5+13.8
    3. 3Claude Sonnet 5.5+12.5

    29 of 41 models · Arena · Oct 2

  • Independent test

    Finance Agent v2

    Multi-step financial analyst tasks that require research, calculation and judgement.

    1. 1Gemini 4 Argon65.4%
    2. 2Gemini 3.8 Flash61.4%
    3. 3Muse Spark 1.360.0%

    26 of 41 models · Vals AI · Oct 7

  • Independent test

    Excel Modeling Benchmark

    Building and editing complex financial models in spreadsheets.

    1. 1Claude Fable 5.176.7%
    2. 2Claude Opus 5.575.9%
    3. 3Claude Sonnet 5.575.7%

    26 of 41 models · Vals AI · Oct 7

  • Independent test

    Legal Research Bench

    Answering questions of U.S. law through case and statute research, with citations that have to hold up.

    1. 1Claude Fable 5.155.3%
    2. 1Muse Spark 1.355.3%
    3. 3Gemini 4 Argon54.8%

    26 of 41 models · Vals AI · Oct 7

  • Independent test

    Tax Agent Bench

    Research-grade U.S. corporate tax questions answered with cited authority.

    1. 1Claude Fable 5.177.6%
    2. 2Gemini 4 Argon76.2%
    3. 3Claude Sonnet 5.573.4%

    26 of 41 models · Vals AI · Oct 7

Reasoning & science

Hard problems with no shortcut: research-level science, novel puzzles and abstract reasoning.

  • Independent test

    Humanity's Last Exam

    2,158 expert-written questions at the frontier of academic knowledge across maths, sciences and humanities.

    1. 1Claude Opus 5.561.4%
    2. 2Claude Fable 5.159.1%
    3. 3Gemini 4 Argon57.1%

    39 of 41 models · Artificial Analysis · Oct 8

  • Independent test

    CritPt

    Seventy unpublished, research-level physics problems graded by an official server.

    1. 1Claude Opus 5.531.7%
    2. 1GPT-6 Astra31.7%
    3. 1GPT-6.1 Sol31.7%

    39 of 41 models · Artificial Analysis · Oct 8

  • Independent testSaturating

    ARC-AGI-2

    Abstract visual puzzles built to be easy for people and unfamiliar to models, verified on a semi-private set.

    1. 1GPT-6 Astra95.0%
    2. 2GPT-6.1 Sol94.2%
    3. 3Claude Opus 5.593.3%

    16 of 41 models · ARC Prize · Oct 8

  • Independent test

    ARC-AGI-3

    Interactive, game-like environments with no instructions: the agent has to explore, infer the goal and learn the rules as it plays. 100% means beating every game as efficiently as a person.

    1. 1GPT-6 Astra62.7%
    2. 2GPT-6.1 Sol52.7%
    3. 3Gemini 3.8 Flash10.4%

    6 of 41 models · ARC Prize · Oct 8

  • Independent testSaturating

    GPQA Diamond

    Graduate-level multiple-choice questions in physics, chemistry and biology that resist simple web search.

    1. 1GPT-6 Astra96.1%
    2. 2Gemini 3.8 Flash95.3%
    3. 3Gemini 3.1 Pro94.1%

    27 of 41 models · Artificial Analysis · Oct 8

Knowledge & reliability

Does it know the answer, admit when it does not, and keep track across very long inputs?

  • Independent test

    AA-Omniscience

    6,000 factual questions across professional domains, scoring both accuracy and how often the model invents an answer instead of abstaining.

    1. 1Claude Opus 5.546
    2. 2Claude Fable 5.143
    3. 2GPT-6 Astra43

    39 of 41 models · Artificial Analysis · Oct 8

  • Independent test

    AA-LCR

    Reasoning questions (v1.1) whose answers require combining information spread across very long document sets.

    1. 1Kimi K388.7%
    2. 2Step 5 Preview88.3%
    3. 3MiMo-V2.6-Pro86.3%

    39 of 41 models · Artificial Analysis · Oct 8

Vision

Understanding images, charts, screenshots and documents.

  • Independent test

    MMMU-Pro

    1,730 college-level questions that can only be answered by reading an image — charts, diagrams, figures — with ten answer options.

    1. 1Claude Opus 5.587.7%
    2. 2GPT-6 Astra86.9%
    3. 3GPT-6.1 Sol86.0%

    27 of 41 models · Artificial Analysis · Oct 8

  • Human preference

    Arena · Vision

    Blind head-to-head votes on answers to prompts that include images.

    1. 1Qwen3.8-Max1301
    2. 2GPT-6.1 Sol1291
    3. 3Gemini 3.8 Flash1290

    19 of 41 models · Arena · Oct 2

Model Index

An independent snapshot of frontier SI models and how they perform on third-party benchmarks.

Data as of Oct 8, 2026

Explore

  • All models
  • Compare
  • Benchmarks
  • Model finder
  • Labs
  • Methodology & sources

Sources

Benchmark results come from Artificial Analysis, Arena, Vals AI and ARC Prize. Specifications and prices come from each lab's own documentation.

Not affiliated with any lab or evaluator. Names and logos are trademarks of their respective owners and are used only to identify them. Scores change often — check the linked source before making a decision that depends on one.

  • Overview
  • Models
  • Compare
  • Benchmarks
  • Finder
  • Overview
  • Models
  • Compare
  • Benchmarks
  • Finder
  • Overview
  • Models
  • Compare
  • Benchmarks
  • Finder