Benchmark leaderboard - 50 entries on 2 boards
ARC-AGI leaderboard
ARC-AGI-2 is designed to stress-test the capabilities of state-of-the-art AI reasoning systems, and ARC-AGI-3 is an interactive reasoning benchmark which challenges AI agents to explore novel environments, acquire goals on the fly, build adaptable world models, and learn continuously (arcprize.org). As of October 17, 2026, GPT-6 Astra (Max) ranks first on the ARC-AGI-2 board with 95.0% and GPT-6 Astra (High) ranks first on the ARC-AGI-3 board with 99.9%.
Last verified from arcprize.org. Read the caveats
This page covers 2 separate boards: ARC-AGI-2 and ARC-AGI-3. Each is scored on its own terms, so compare rows only within the same board.
ARC-AGI-2
25 entries, in the board's order
| 1 | GPT-6 Astra (Max) Rank 1 on ARC-AGI-2; reasoning level: Max; cost $1.12/task; date is the leaderboard Release Date column Official leaderboard | 95.0% |
| 2 | GPT-6.1 Sol (Max) Rank 2 on ARC-AGI-2; reasoning level: Max; cost $0.254/task; date is the leaderboard Release Date column Official leaderboard | 94.2% |
| 3 | Claude Opus 5.5 (High) Rank 3 on ARC-AGI-2; reasoning level: High; cost $0.408/task; date is the leaderboard Release Date column Official leaderboard | 93.3% |
| 4 | GPT-6 Astra (XHigh) Rank 4 on ARC-AGI-2; reasoning level: XHigh; cost $0.829/task; date is the leaderboard Release Date column Official leaderboard | 93.3% |
| 5 | Claude Opus 5.5 (XHigh) Rank 5 on ARC-AGI-2; reasoning level: XHigh; cost $0.671/task; date is the leaderboard Release Date column Official leaderboard | 92.5% |
| 6 | GPT-5.6 Sol (Max) Rank 6 on ARC-AGI-2; reasoning level: Max; cost $1.44/task; date is the leaderboard Release Date column Official leaderboard | 92.5% |
| 7 | GPT-6 Astra (High) Rank 7 on ARC-AGI-2; reasoning level: High; cost $0.668/task; date is the leaderboard Release Date column Official leaderboard | 92.1% |
| 8 | GPT-6 Astra (Medium) Rank 8 on ARC-AGI-2; reasoning level: Medium; cost $0.480/task; date is the leaderboard Release Date column Official leaderboard | 92.1% |
| 9 | GPT-6.1 Sol (XHigh) Rank 9 on ARC-AGI-2; reasoning level: XHigh; cost $0.178/task; date is the leaderboard Release Date column Official leaderboard | 91.7% |
| 10 | GPT-6.1 Sol (High) Rank 10 on ARC-AGI-2; reasoning level: High; cost $0.135/task; date is the leaderboard Release Date column Official leaderboard | 91.7% |
| 11 | Claude Opus 5.5 (Max) Rank 11 on ARC-AGI-2; reasoning level: Max; cost $1.85/task; date is the leaderboard Release Date column Official leaderboard | 91.7% |
| 12 | Gemini 4 Argon (High) Rank 12 on ARC-AGI-2; reasoning level: High; cost $0.325/task; date is the leaderboard Release Date column Official leaderboard | 91.3% |
| 13 | Claude Opus 5 (Max) Rank 13 on ARC-AGI-2; reasoning level: Max; cost $2.06/task; date is the leaderboard Release Date column Official leaderboard | 90.4% |
| 14 | Claude Fable 5.1 (Max) Rank 14 on ARC-AGI-2; reasoning level: Max; cost $4.49/task; date is the leaderboard Release Date column Official leaderboard | 90.0% |
| 15 | Claude Fable 5.1 (XHigh) Rank 15 on ARC-AGI-2; reasoning level: XHigh; cost $3.12/task; date is the leaderboard Release Date column Official leaderboard | 90.0% |
| 16 | GPT-5.6 Sol (XHigh) Rank 16 on ARC-AGI-2; reasoning level: XHigh; cost $1.04/task; date is the leaderboard Release Date column Official leaderboard | 90.0% |
| 17 | GPT-6 Sol (Max) Rank 17 on ARC-AGI-2; reasoning level: Max; cost $0.439/task; date is the leaderboard Release Date column Official leaderboard | 89.6% |
| 18 | Gemini 3.8 Flash (High) Rank 18 on ARC-AGI-2; reasoning level: High; cost $0.400/task; date is the leaderboard Release Date column Official leaderboard | 89.2% |
| 19 | Claude Fable 5 (Max) Rank 19 on ARC-AGI-2; reasoning level: Max; cost $5.45/task; date is the leaderboard Release Date column Official leaderboard | 89.2% |
| 20 | Claude Fable 5.1 (High) Rank 20 on ARC-AGI-2; reasoning level: High; cost $1.67/task; date is the leaderboard Release Date column Official leaderboard | 88.8% |
| 21 | Claude Opus 5 (High) Rank 21 on ARC-AGI-2; reasoning level: High; cost $1.45/task; date is the leaderboard Release Date column Official leaderboard | 88.3% |
| 22 | Claude Fable 5 (XHigh) Rank 22 on ARC-AGI-2; reasoning level: XHigh; cost $3.00/task; date is the leaderboard Release Date column Official leaderboard | 88.3% |
| 23 | Claude Opus 5.5 (Medium) Rank 23 on ARC-AGI-2; reasoning level: Medium; cost $0.337/task; date is the leaderboard Release Date column Official leaderboard | 87.5% |
| 24 | Claude Fable 5 (High) Rank 24 on ARC-AGI-2; reasoning level: High; cost $2.44/task; date is the leaderboard Release Date column Official leaderboard | 87.5% |
| 25 | Gemini 4 Argon (Medium) Rank 25 on ARC-AGI-2; reasoning level: Medium; cost $0.137/task; date is the leaderboard Release Date column Official leaderboard | 87.1% |
ARC-AGI-3
25 entries, in the board's order
| 1 | GPT-6 Astra (High) Rank 1 on ARC-AGI-3; Provider Adapter harness; reasoning level: High; cost $18.8K; date is the leaderboard Release Date column Official leaderboard | 99.9% |
| 2 | GPT-6 Astra (Max) Rank 2 on ARC-AGI-3; Provider Adapter harness; reasoning level: Max; cost $17.3K; date is the leaderboard Release Date column Official leaderboard | 98.6% |
| 3 | GPT-6 Astra (XHigh) Rank 3 on ARC-AGI-3; Provider Adapter harness; reasoning level: XHigh; cost $18.1K; date is the leaderboard Release Date column Official leaderboard | 98.4% |
| 4 | GPT-6 Astra (Medium) Rank 4 on ARC-AGI-3; Provider Adapter harness; reasoning level: Medium; cost $19.3K; date is the leaderboard Release Date column Official leaderboard | 98.4% |
| 5 | GPT-6 Astra (Low) Rank 5 on ARC-AGI-3; Provider Adapter harness; reasoning level: Low; cost $21.3K; date is the leaderboard Release Date column Official leaderboard | 98.0% |
| 6 | Gemini 4 Argon (Low) Rank 6 on ARC-AGI-3; Provider Adapter harness; reasoning level: Low; cost $4.8K; date is the leaderboard Release Date column Official leaderboard | 97.5% |
| 7 | Gemini 4 Argon (High) Rank 7 on ARC-AGI-3; Provider Adapter harness; reasoning level: High; cost $4.6K; date is the leaderboard Release Date column Official leaderboard | 97.3% |
| 8 | Gemini 4 Argon (Medium) Rank 8 on ARC-AGI-3; Provider Adapter harness; reasoning level: Medium; cost $5.0K; date is the leaderboard Release Date column Official leaderboard | 97.3% |
| 9 | GPT-6 Astra (None) Rank 9 on ARC-AGI-3; Provider Adapter harness; reasoning level: None; cost $23.5K; date is the leaderboard Release Date column Official leaderboard | 96.7% |
| 10 | GPT-6.1 Sol (XHigh) Rank 10 on ARC-AGI-3; Provider Adapter harness; reasoning level: XHigh; cost $4.4K; date is the leaderboard Release Date column Official leaderboard | 96.4% |
| 11 | GPT-6.1 Sol (Max) Rank 11 on ARC-AGI-3; Provider Adapter harness; reasoning level: Max; cost $3.8K; date is the leaderboard Release Date column Official leaderboard | 96.2% |
| 12 | GPT-6.1 Sol (High) Rank 12 on ARC-AGI-3; Provider Adapter harness; reasoning level: High; cost $5.4K; date is the leaderboard Release Date column Official leaderboard | 95.0% |
| 13 | GPT-6.1 Sol (Medium) Rank 13 on ARC-AGI-3; Provider Adapter harness; reasoning level: Medium; cost $5.8K; date is the leaderboard Release Date column Official leaderboard | 91.0% |
| 14 | GPT-6.1 Sol (Low) Rank 14 on ARC-AGI-3; Provider Adapter harness; reasoning level: Low; cost $6.6K; date is the leaderboard Release Date column Official leaderboard | 82.8% |
| 15 | GPT-6 Astra (Max) Rank 15 on ARC-AGI-3; Standard harness; reasoning level: Max; cost $26.1K; date is the leaderboard Release Date column Official leaderboard | 62.7% |
| 16 | GPT-6 Astra (XHigh) Rank 16 on ARC-AGI-3; Standard harness; reasoning level: XHigh; cost $37.3K; date is the leaderboard Release Date column Official leaderboard | 59.3% |
| 17 | GPT-6 Astra (High) Rank 17 on ARC-AGI-3; Standard harness; reasoning level: High; cost $40.7K; date is the leaderboard Release Date column Official leaderboard | 54.8% |
| 18 | GPT-6.1 Sol (Max) Rank 18 on ARC-AGI-3; Standard harness; reasoning level: Max; cost $7.6K; date is the leaderboard Release Date column Official leaderboard | 52.7% |
| 19 | Gemini 4 Argon (Low) Rank 19 on ARC-AGI-3; Standard harness; reasoning level: Low; cost $5.8K; date is the leaderboard Release Date column Official leaderboard | 49.8% |
| 20 | Gemini 4 Argon (Medium) Rank 20 on ARC-AGI-3; Standard harness; reasoning level: Medium; cost $6.2K; date is the leaderboard Release Date column Official leaderboard | 46.3% |
| 21 | GPT-6.1 Sol (XHigh) Rank 21 on ARC-AGI-3; Standard harness; reasoning level: XHigh; cost $9.1K; date is the leaderboard Release Date column Official leaderboard | 39.9% |
| 22 | GPT-6 Astra (Medium) Rank 22 on ARC-AGI-3; Standard harness; reasoning level: Medium; cost $48.1K; date is the leaderboard Release Date column Official leaderboard | 38.6% |
| 23 | GPT-6 Astra (None) Rank 23 on ARC-AGI-3; Standard harness; reasoning level: None; cost $49.8K; date is the leaderboard Release Date column Official leaderboard | 35.2% |
| 24 | Gemini 3.8 Flash (High) Rank 24 on ARC-AGI-3; Provider Adapter harness; reasoning level: High; cost $4.5K; date is the leaderboard Release Date column Official leaderboard | 35.0% |
| 25 | Claude Opus 5 (High) Rank 25 on ARC-AGI-3; Standard harness; reasoning level: High; cost $20.7K; date is the leaderboard Release Date column Official leaderboard | 30.2% |
About this board
- Separate eval or verification dates per entry: the arcprize.org/leaderboard table lists only each model's Release Date, so the date column shows that release date.
- ARC-AGI-3 results for Claude Opus 5.5 and Claude Fable 5/5.1: shown as N/A on the leaderboard.
- Non-model rows left out of the rankings: Human Panel (ARC-AGI-2 100.0%, $17.00/task), Avg. Mturker and Stem Grad.
- ARC-AGI-1 not ranked: it is an older version, and this page covers the current versions (ARC-AGI-2 and ARC-AGI-3)
FAQ
What does ARC-AGI measure?
ARC-AGI-2 is designed to stress-test the capabilities of state-of-the-art AI reasoning systems, and ARC-AGI-3 is an interactive reasoning benchmark which challenges AI agents to explore novel environments, acquire goals on the fly, build adaptable world models, and learn continuously (arcprize.org). This page covers 2 separate boards: ARC-AGI-2 and ARC-AGI-3. Each is scored on its own terms, so compare rows only within the same board.
Which model leads ARC-AGI?
As of October 17, 2026, GPT-6 Astra (Max) ranks first on the ARC-AGI-2 board with 95.0% and GPT-6 Astra (High) ranks first on the ARC-AGI-3 board with 99.9%. That is the board's own order, not a Developers Digest ranking.
Where do these ARC-AGI scores come from?
All 50 entries come from the official leaderboard at arcprize.org. Every entry links to its source, which was read on October 17, 2026, and lists the setting it was run under.
Related reading
- OpenAI's Efficiency Ledger: Serving Costs Down 20%, ARC-AGI-3 Up 3x With No Model ChangeThe OpenAI "Building abundant intelligence" essay puts numbers on efficiency: 20% lower serving costs and a 3x ARC-AGI-3 jump from two settings alone.
- GPT-6 Astra Release Guide: Benchmarks, $10/$50 Pricing, and How to Run It in Codex and OpenCodeGPT-6 Astra release guide: OpenAI's max-capability model, its $10/$50 pricing and 1.05M context, what the benchmarks do and do not say, and commands to try it.
- GPT-6 in 7 Minutes: Astra, Sol, and Luna Explained for DevelopersGPT-6 in 7 minutes: what Astra, Sol and Luna are, what each tier costs per million tokens, the new API features, and how to try each one in Codex and the API.
- Kill Your Agent Runs EarlyKill your agent runs early: production traces show context goes stale at turn boundaries, and the fixes that moved numbers are kill, carry the state, restart.
- Prime Agent: A Self-Improving Coding Harness Where Everything Is PythonPrime Agent is Prime Intellect's open-source coding harness with one tool, a persistent IPython kernel. How it works, its benchmarks, and Claude Code compared.
- Claude Opus 5: Near-Fable Intelligence at Half the CostClaude Opus 5 benchmarks, pricing and HN reaction: near Fable 5 performance at the Opus 4.8 price, and what it means for choosing a daily-driver model.
Sources
Last verified . Each row links to the page its score was copied from.
Newsletter
Get new benchmark results as they land
One email per week. New model releases, the leaderboard moves that matter and what they mean for what you build.