Last checked: 17 November 2026, 00:48 UTC. ARC-AGI-3 tests whether AI systems can learn unfamiliar interactive games the way people do. We track the best published score for each frontier model on ARC Prize's semi-private set, checked against ARC Prize's own data every few hours. Scores are the share of games solved. ARC Prize 2026 results are due on 4 December 2026.
With a provider adapter
The model runs inside its provider's own agent harness for ARC-AGI-3 games. Best published configuration per model.
| # | Model | Lab | Effort | Score | Run cost | Released |
|---|---|---|---|---|---|---|
| 1 | GPT-6 Astra# | OpenAI | High | 99.9% | $18,817 | 2 Sep 2026 |
| 2 | Gemini 4 Carbon# | Medium | 99.1% | $4,560 | 10 Nov 2026 | |
| 3 | Claude Fable 5.2# | Anthropic | High | 97.7% | $18,625 | 9 Nov 2026 |
| 4 | Gemini 4 Argon# | Low | 97.5% | $4,770 | 13 Oct 2026 | |
| 5 | GPT-6.1 Sol# | OpenAI | XHigh | 96.4% | $4,360 | 29 Sep 2026 |
| 6 | Claude Sonnet 5.5# | Anthropic | XHigh | 72.4% | $8,170 | 28 Sep 2026 |
| 7 | Gemini 3.8 Flash# | High | 35.0% | $4,522 | 2 Sep 2026 | |
| 8 | GPT-6 Sol# | OpenAI | Max | 23.0% | $8,722 | 22 Sep 2026 |
| 9 | Grok 4.7# | xAI | High | 10.0% | $4,774 | 21 Sep 2026 |
| 10 | DeepSeek V4.1 Flash 1017# | DeepSeek | Max | 5.8% | $1,179 | 17 Oct 2026 |
Model alone
The same benchmark with no provider harness: the model plays each game through ARC Prize's standard interface. Best published configuration per model.
| # | Model | Lab | Effort | Score | Run cost | Released |
|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.2# | Anthropic | XHigh | 81.2% | $21,949 | 9 Nov 2026 |
| 2 | Gemini 4 Carbon# | Low | 65.1% | $7,244 | 10 Nov 2026 | |
| 3 | GPT-6 Astra# | OpenAI | Max | 62.7% | $26,098 | 2 Sep 2026 |
| 4 | GPT-6.1 Sol# | OpenAI | Max | 52.7% | $7,632 | 29 Sep 2026 |
| 5 | Gemini 4 Argon# | Low | 49.8% | $5,821 | 13 Oct 2026 | |
| 6 | Claude Opus 5# | Anthropic | High | 30.2% | $20,657 | 24 Jul 2026 |
| 7 | Claude Sonnet 5.5# | Anthropic | XHigh | 24.3% | $7,394 | 28 Sep 2026 |
| 8 | Gemini 3.8 Flash# | High | 10.4% | $4,398 | 2 Sep 2026 | |
| 9 | GPT-5.6 Sol# | OpenAI | Max | 7.8% | $25,064 | 9 Jul 2026 |
| 10 | GPT-6 Sol# | OpenAI | Max | 4.6% | $5,554 | 22 Sep 2026 |
| 11 | Grok 4.6# | xAI | XHigh | 2.1% | $5,612 | 11 Aug 2026 |
| 12 | Claude Opus 4.6# | Anthropic | Max | 0.5% | $8,866 | 17 Dec 2025 |
Cite this page
KernelPost. "ARC-AGI-3 leaderboard: how frontier models score." Last checked 17 November 2026. https://kernelpost.com/arc-agi-3-leaderboard/
Hover a row for its # link to cite a single row.