Skip to content
Bencher

TestsBenchmarks · Reasoning

ARC-AGI-3

Score on interactive, game-like ARC-AGI-3 environments that require exploration and learning; vendor "provider adapter" harnesses score far higher than bare models.

% solvedHigher is betterThe test's websiteOfficial page ↗

The best 11 · 11 models tested

Top 11 · best setting per model · 11 models, 39 results (33 independent)

Measured by independent testersIndependent Reported by the makerVendor-reported
ARC-AGI-3 leaderboard
Thinking levelSetting
1GPT-6 AstraOpenAI99.9%HighGPT-6 Astra - Provider Adapter (High)Independent testIndependentARC Prize ↗Provider Adapter (vendor-built agent harness)30 Sep 2026
2GPT-6.1 SolOpenAI96.4%Extra highGPT-6.1 Sol - Provider Adapter (XHigh)Independent testIndependentARC Prize ↗Provider Adapter (vendor-built agent harness)30 Sep 2026
3Gemini 3.8 FlashGoogle (Gemini / DeepMind)35.0%HighGemini 3.8 Flash - Provider Adapter (High)Independent testIndependentARC Prize ↗Provider Adapter (vendor-built agent harness)30 Sep 2026
4Claude Opus 5Anthropic30.2%HighClaude Opus 5 (High)Independent testIndependentARC Prize ↗—30 Sep 2026
5GPT-6 SolOpenAI23.0%MaxGPT-6 Sol - Provider Adapter (Max)Independent testIndependentARC Prize ↗Provider Adapter (vendor-built agent harness)30 Sep 2026
6GPT-5.6 SolOpenAI7.8%MaxGPT-5.6 Sol (Max)Independent testIndependentARC Prize ↗—30 Sep 2026
7Claude Opus 4.8Anthropic1.5%Highadaptive thinking, effort=highMaker's own figureVendor-reportedAnthropic ↗—24 Jul 2026
8GPT-5.6 TerraOpenAI0.8%MaxGPT-5.6 Terra (Max)Independent testIndependentARC Prize ↗—30 Sep 2026
9GPT-6 LunaOpenAI0.6%MaxGPT-6 Luna - Provider Adapter (Max)Independent testIndependentARC Prize ↗Provider Adapter (vendor-built agent harness)30 Sep 2026
10GPT-5.5OpenAI0.4%DefaultGPT-5.6 launch table (effort not stated; text cites max for headline results)Maker's own figureVendor-reportedOpenAI ↗—9 Jul 2026
11GPT-5.6 LunaOpenAI0.2%DefaultGPT-5.6 launch table (effort not stated; text cites max for headline results)Maker's own figureVendor-reportedOpenAI ↗—9 Jul 2026