Skip to content
Yardstick
  1. Benchmarks
  2. ARC-AGI-3

ARC-AGI-3

Unfamiliar turn-based games with no instructions: explore, work out the goal, then win efficiently.

ReasoningRun by ARC Prize FoundationOfficial leaderboard (opens in a new tab)

Top score
62.7%GPT-6 Astra
Median of field
10.4%5 models scored
Human baseline
100%Median human action efficiency
Data as of
Oct 8, 2026ARC Prize Verified leaderboard, Standard harness

Standings

Bars share one axis from 0% to 100%.

  1. Rank 1
    GPT-6 Astra62.7%

    OpenAI · $10 / $50New

  2. Rank 2
    GPT-6.1 Sol52.7%

    OpenAI · $2 / $10New

  3. Rank 3
    Gemini 3.8 Flash10.4%

    Google · $0.75 / $3.75New

  4. Rank 4
    Grok 4.71.8%

    SpaceXAI · $2 / $6New

  5. Rank 5
    GPT-6 Luna0.2%

    OpenAI · $0.10 / $0.50New

Not reported for ARC-AGI-3 · 24

  • ·
    Claude Fable 5.1—Not reported

    Anthropic · $10 / $50New

  • ·
    Claude Haiku 5.5—Not reported

    Anthropic · $0.10 / $0.50New

  • ·
    Claude Opus 5.5—Not reported

    Anthropic · $4 / $20New

  • ·
    Claude Sonnet 5.5—Not reported

    Anthropic · $2 / $10New

  • ·
    DeepSeek-V4-Pro—Not reported

    DeepSeek · $1.32 / $3.96NewOpen weights

  • ·
    DeepSeek-V4.1-Flash—Not reported

    DeepSeek · $0.30 / $1.20NewOpen weights

  • ·
    Gemini 3.1 Pro—Not reported

    Google · $2 / $12Preview

  • ·
    Gemini 3.5 Flash-Lite—Not reported

    Google · $0.30 / $2.50

  • ·
    Gemini 4 Argon—Not reported

    Google · $2 / $10NewPreview

  • ·
    Gemma 4 31B—Not reported

    Google · Self-hostedOpen weights

  • ·
    GLM-5.3—Not reported

    Z.ai · $1.40 / $4.40NewOpen weights

  • ·
    GLM-5.3-Flash—Not reported

    Z.ai · $0.15 / $0.50NewOpen weights

  • ·
    Kimi K2.7 Code—Not reported

    Moonshot AI · $0.95 / $4Open weights

  • ·
    Kimi K3—Not reported

    Moonshot AI · $3 / $15Open weights

  • ·
    MiMo-V2.6-Pro—Not reported

    Xiaomi · $0.435 / $0.87NewOpen weights

  • ·
    MiniMax M3—Not reported

    MiniMax · $0.30 / $1.20Open weights

  • ·
    Mistral Large 4—Not reported

    Mistral AI · $1.36 / $4.18NewPreview

  • ·
    Mistral Medium 3.5—Not reported

    Mistral AI · $1.50 / $7.50Open weights

  • ·
    Mistral Small 4—Not reported

    Mistral AI · $0.15 / $0.60Open weights

  • ·
    Muse Glimmer 30B—Not reported

    Meta · Self-hostedNewOpen weights

  • ·
    Muse Spark 1.3—Not reported

    Meta · $1.25 / $4.25New

  • ·
    Qwen3.8-27B—Not reported

    Alibaba · $0.50 / $3NewOpen weights

  • ·
    Qwen3.8-Flash-Next—Not reported

    Alibaba · $0.15 / $0.47NewOpen weights

  • ·
    Qwen3.8-Max—Not reported

    Alibaba · $2 / $6

What it measures

Interactive reasoning. An agent is dropped into a novel game environment and must discover the rules and the objective by acting, then plan its way through.

How it is scored

The score is normalised to human action efficiency: 100% means beating every game in as few moves as the median human tester. Shown here for ARC Prize's Standard harness only, so rows are like for like.

Read it with care

The harness changes everything. With ARC Prize's Provider Adapter harness, which preserves a provider's reasoning state between turns, GPT-6 Astra reaches 99.95% against 62.71% on the Standard harness. Only a handful of models have been run.

Other benchmarks

  • Artificial Analysis Intelligence Index
  • Epoch Capabilities Index
  • LiveBench
  • Terminal-Bench 4.0
  • SWE-Bench Pro (V2 Hard)
  • Arena WebDev
  • GDPval-AA
  • Vending-Bench 2
  • ARC-AGI-2
  • FrontierMath (Tiers 1–3)
  • MathArena
  • SimpleBench
  • Humanity's Last Exam (Diamond)
  • GPQA Diamond
  • Arena Vision
  • Arena Text
Yardstick

A plain reading of how today’s SI models compare. Scores are taken from each benchmark’s own leaderboard wherever one exists; anything a lab reported about its own model is marked.

Data as of Oct 8, 2026.

Explore

  • Models
  • Compare
  • Benchmarks
  • Value
  • Labs
  • Methodology & sources

Sources

  • Artificial Analysis (opens in a new tab)
  • Arena (opens in a new tab)
  • Epoch AI Benchmarking Hub (opens in a new tab)
  • ARC Prize (opens in a new tab)
  • Terminal-Bench (opens in a new tab)
  • Scale AI Labs (opens in a new tab)

Yardstick is an independent reference and is not affiliated with, sponsored by or endorsed by any lab listed. Product names and logos are trademarks of their respective owners and are used only to identify them. Benchmarks are imperfect proxies: check a model on your own task before you commit to it.

  • Models
  • Compare
  • Benchmarks
  • Value
  • Labs
  • Models
  • Compare
  • Benchmarks
  • Value
  • Labs
  • Models
  • Compare
  • Benchmarks
  • Value
  • Labs