Abstraction and Reasoning Corpus for AGI v3 (ARC-AGI-3)
We show this table for reference; we do not rank on it.
An interactive successor to ARC-AGI-2 that evaluates whether an AI agent can learn unfamiliar task mechanics through action and feedback.
Earlier track
ARC-AGI-2 evaluates static grid-puzzle reasoning. It remains a separate benchmark track, and its scores are not directly comparable with ARC-AGI-3.
View ARC-AGI-2Semi-Private score on ARC-AGI-3 — Captured 2026-10-25
We mirror the published semi-private score view for ARC-AGI-3. GPT-6 Astra (Max) leads the public snapshot at 62.7%, followed by GPT-6 Astra (XHigh) (59.3%) and GPT-6 Astra (High) (54.8%). We do not use these results to rank models overall.
GPT-6 Astra (Max)
OpenAI
Standard · max reasoning
GPT-6 Astra (XHigh)
OpenAI
Standard · xhigh reasoning
GPT-6 Astra (High)
OpenAI
Standard · high reasoning
68 configurationsReasoning15% of category scoreCurrentUpdated Captured 2026-10-25
Semi-Private score table (68 configurations)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | GPT-6 Astra (Max)OpenAIStandard · max reasoning | 62.7% | Not reported | Closed |
| 2 | GPT-6 Astra (XHigh)OpenAIStandard · xhigh reasoning | 59.3% | Not reported | Closed |
| 3 | GPT-6 Astra (High)OpenAIStandard · high reasoning | 54.8% | Not reported | Closed |
| 4 | GPT-6.1 Sol (Max)OpenAIStandard · max reasoning | 52.7% | Not reported | Closed |
| 5 | Gemini 4 Argon (Low)GoogleStandard · low reasoning | 49.8% | Not reported | Closed |
| 6 | Gemini 4 Argon (Medium)GoogleStandard · medium reasoning | 46.3% | Not reported | Closed |
| 7 | GPT-6.1 Sol (XHigh)OpenAIStandard · xhigh reasoning | 39.9% | Not reported | Closed |
| 8 | GPT-6 Astra (Medium)OpenAIStandard · medium reasoning | 38.6% | Not reported | Closed |
| 9 | GPT-6 Astra (None)OpenAIStandard · none reasoning | 35.2% | Not reported | Closed |
| 10 | Claude Opus 5 (High)AnthropicStandard · high reasoning | 30.2% | Not reported | Closed |
| 11 | GPT-6.1 Sol (High)OpenAIStandard · high reasoning | 26.7% | Not reported | Closed |
| 12 | Claude Sonnet 5.5 (XHigh)AnthropicStandard · xhigh reasoning | 24.3% | Not reported | Unknown |
| 13 | GPT-6 Astra (Low)OpenAIStandard · low reasoning | 17.5% | Not reported | Closed |
| 14 | GPT-6.1 Sol (Medium)OpenAIStandard · medium reasoning | 10.6% | Not reported | Closed |
| 15 | Gemini 3.8 Flash (High)GoogleStandard · high reasoning | 10.4% | Not reported | Closed |
| 16 | GPT-5.6 Sol (Max)OpenAIStandard · max reasoning | 7.8% | Not reported | Closed |
| 17 | GPT-5.6 Sol (XHigh)OpenAIStandard · xhigh reasoning | 7.0% | Not reported | Closed |
| 18 | Gemini 3.8 Flash (Low)GoogleStandard · low reasoning | 6.0% | Not reported | Closed |
| 19 | GPT-6 Sol (Max)OpenAIStandard · max reasoning | 4.6% | Not reported | Closed |
| 20 | Gemini 3.8 Flash (Medium)GoogleStandard · medium reasoning | 4.0% | Not reported | Closed |
| 21 | GPT-6.1 Sol (Low)OpenAIStandard · low reasoning | 3.9% | Not reported | Closed |
| 22 | GPT-5.6 Sol (High)OpenAIStandard · high reasoning | 2.1% | Not reported | Closed |
| 23 | Grok 4.6 (XHigh)xAIStandard · xhigh reasoning | 2.1% | Not reported | Closed |
| 24 | Grok 4.7 (XHigh)xAIStandard · xhigh reasoning | 1.8% | Not reported | Closed |
| 25 | GPT-6 Sol (XHigh)OpenAIStandard · xhigh reasoning | 1.8% | Not reported | Closed |
| 26 | Grok 4.7 (High)xAIStandard · high reasoning | 1.7% | Not reported | Closed |
| 27 | Grok 4.7 (Medium)xAIStandard · medium reasoning | 1.7% | Not reported | Closed |
| 28 | Claude Opus 4.8 (High)AnthropicStandard · high reasoning | 1.5% | Not reported | Closed |
| 29 | Claude Sonnet 5.5 (High)AnthropicStandard · high reasoning | 1.2% | Not reported | Unknown |
| 30 | GPT-5.6 Sol (Medium)OpenAIStandard · medium reasoning | 1.1% | Not reported | Closed |
| 31 | GPT-6 Sol (High)OpenAIStandard · high reasoning | 0.9% | Not reported | Closed |
| 32 | DeepSeek V4.1 Flash (Max)DeepSeekStandard · max reasoning | 0.9% | Not reported | Unknown |
| 33 | GPT-5.6 Terra (Max)OpenAIStandard · max reasoning | 0.8% | Not reported | Closed |
| 34 | GPT-5.6 Terra (XHigh)OpenAIStandard · xhigh reasoning | 0.7% | Not reported | Closed |
| 35 | Anthropic Opus 4.6 (Max)AnthropicStandard · max reasoning | 0.5% | Not reported | Closed |
| 36 | GPT-5.6 Terra (High)OpenAIStandard · high reasoning | 0.5% | Not reported | Closed |
| 37 | DeepSeek V4.1 Flash (High)DeepSeekStandard · high reasoning | 0.5% | Not reported | Unknown |
| 38 | Claude Haiku 5.5 (Max)AnthropicStandard · max reasoning | 0.5% | Not reported | Closed |
| 39 | GPT-5.5 (High)OpenAIStandard · high reasoning | 0.4% | Not reported | Closed |
| 40 | Gemini 3.1 Pro (Preview)GoogleStandard | 0.4% | Not reported | Closed |
| 41 | GPT-6 Sol (Medium)OpenAIStandard · medium reasoning | 0.4% | Not reported | Closed |
| 42 | Claude Sonnet 5.5 (Low)AnthropicStandard · low reasoning | 0.4% | Not reported | Unknown |
| 43 | Claude Sonnet 5.5 (Medium)AnthropicStandard · medium reasoning | 0.4% | Not reported | Unknown |
| 44 | DeepSeek V4.1 Flash (Low)DeepSeekStandard · low reasoning | 0.3% | Not reported | Unknown |
| 45 | DeepSeek V4.1 Flash (None)DeepSeekStandard · none reasoning | 0.3% | Not reported | Unknown |
| 46 | GPT-5.6 Sol (Low)OpenAIStandard · low reasoning | 0.3% | Not reported | Closed |
| 47 | Grok 4.5 (Medium)xAIStandard · medium reasoning | 0.3% | Not reported | Closed |
| 48 | Grok 4.5 (High)xAIStandard · high reasoning | 0.3% | Not reported | Closed |
| 49 | GPT-6 Sol (None)OpenAIStandard · none reasoning | 0.3% | Not reported | Closed |
| 50 | Grok 4.5 (Low)xAIStandard · low reasoning | 0.3% | Not reported | Closed |
| 51 | Grok 4.7 (Low)xAIStandard · low reasoning | 0.2% | Not reported | Closed |
| 52 | GPT-5.4 (High)OpenAIStandard · high reasoning | 0.2% | Not reported | Closed |
| 53 | GPT-6 Luna (Medium)OpenAIStandard · medium reasoning | 0.2% | Not reported | Closed |
| 54 | GPT-6 Luna (High)OpenAIStandard · high reasoning | 0.2% | Not reported | Closed |
| 55 | GPT-5.6 Luna (Max)OpenAIStandard · max reasoning | 0.2% | Not reported | Closed |
| 56 | Opus 4.7 (High)AnthropicStandard · high reasoning | 0.2% | Not reported | Closed |
| 57 | GPT-5.6 Luna (Low)OpenAIStandard · low reasoning | 0.2% | Not reported | Closed |
| 58 | GPT-5.6 Luna (Medium)OpenAIStandard · medium reasoning | 0.2% | Not reported | Closed |
| 59 | GPT-6 Luna (XHigh)OpenAIStandard · xhigh reasoning | 0.2% | Not reported | Closed |
| 60 | GPT-6 Sol (Low)OpenAIStandard · low reasoning | 0.1% | Not reported | Closed |
| 61 | GPT-6 Luna (Max)OpenAIStandard · max reasoning | 0.1% | Not reported | Closed |
| 62 | GPT-5.6 Luna (High)OpenAIStandard · high reasoning | 0.1% | Not reported | Closed |
| 63 | Grok 4.20 (Beta Reasoning)xAIStandard | 0.1% | Not reported | Unknown |
| 64 | GPT-5.6 Terra (Medium)OpenAIStandard · medium reasoning | 0.1% | Not reported | Closed |
| 65 | GPT-6 Luna (Low)OpenAIStandard · low reasoning | 0.0% | Not reported | Closed |
| 66 | GPT-6 Luna (None)OpenAIStandard · none reasoning | 0.0% | Not reported | Closed |
| 67 | GPT-5.6 Luna (XHigh)OpenAIStandard · xhigh reasoning | 0.0% | Not reported | Closed |
| 68 | GPT-5.6 Terra (Low)OpenAIStandard · low reasoning | 0.0% | Not reported | Closed |
Standard configurations
These are ARC Prize's published Semi-Private results. Every published reasoning-effort configuration stays separate; the table does not combine efforts or evaluation splits.
Standard and Provider Adapter harnesses are separate protocols. The Provider Adapter retains opaque reasoning state and compacts long conversations. A score gap can reflect those system differences.
API issues affected Gemini 4 Argon's Standard runs: 42 of 55 games completed at low effort and 46 at medium. Scores include partial credit across all 55. The high-effort run completed 19 games and has no published score.
Model pages use the highest declared published reasoning effort within each harness, rather than selecting whichever effort scored highest. Hidden or unpublished incomplete runs are not assigned a score.
The published ARC-AGI-3 snapshot places GPT-6 Astra (Max) first at 62.7%. The third row is 7.9 points behind. The broader top-10 range is 32.6 points, so the table still separates the published systems.
68 configurations have been evaluated on ARC-AGI-3. The benchmark falls in the Reasoning category. BenchLM shows the published table for reference; the per-model scores it stores feed the ranking. The Reasoning leaderboard ranks models by a weighted category score, and ARC-AGI-3 contributes 15% of it. BenchAlign v5.8 also uses it in the overall ranking.
About ARC-AGI-3
Year
2026
Tasks
Interactive game-like tasks with hidden rules
Format
Agentic task completion under a capped evaluation budget
Difficulty
Frontier agentic reasoning
ARC-AGI-3 measures interactive reasoning rather than static grid-puzzle completion. ARC Prize reports Standard and Provider Adapter harness runs separately. The ARC-AGI-3 field stores the Standard harness result. Provider Adapter runs have a separate catalog page and score field. Both pages retain every published effort configuration.
Freshness and provenance
Version
ARC-AGI 3
Refresh cadence
Static
Staleness state
Current
Question availability
Private interactive tasks with public aggregate results
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does ARC-AGI-3 measure?
An interactive successor to ARC-AGI-2 that evaluates whether an AI agent can learn unfamiliar task mechanics through action and feedback.
Which model leads the published ARC-AGI-3 snapshot?
GPT-6 Astra (Max) currently leads the published ARC-AGI-3 snapshot with 62.7% semi-private score. BenchLM shows the published table for reference; the per-model scores it stores feed the ranking.
How many models are evaluated on ARC-AGI-3?
The Captured 2026-10-25 snapshot contains 68 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.