ARC-AGI-2
Das Ergebnis jedes Modells, so, wie der Benchmarking Hub von Epoch AI es veröffentlicht, von Epoch gemessen oder aus der genannten Quelle gesammelt. Facthory hat sie nicht selbst gemessen.
Daten: Epoch AI, „Capabilities & benchmarking“, epoch.ai/benchmarks, lizenziert unter CC BY 4.0. Gelesen am 16.11.2026.
Das Diagramm zeigt die 25 besten, die Tabelle die 100 besten von 275 Ergebnissen.
| # | Modell | Organisation | Ergebnis | Nachweis |
|---|---|---|---|---|
| 1 | GPT-6 Astra (max) API access | OpenAI | 95 % | – |
| 2 | Gemini 4 Carbon (medium) API access | Google DeepMind | 94,2 % | – |
| 3 | GPT-6.1 Sol (max) API access | OpenAI | 94,2 % | – |
| 4 | Claude Fable 5.2 (xhigh) API access | Anthropic | 93,8 % | – |
| 5 | Claude Opus 5.5 (high) API access | Anthropic | 93,3 % | – |
| 6 | GPT-6 Astra (xhigh) API access | OpenAI | 93,3 % | – |
| 7 | Claude Fable 5.2 (max) API access | Anthropic | 92,5 % | – |
| 8 | Claude Opus 5.5 (xhigh) API access | Anthropic | 92,5 % | – |
| 9 | GPT-5.6 Sol (max) API access | OpenAI | 92,5 % | – |
| 10 | Gemini 4 Carbon (high) API access | Google DeepMind | 92,1 % | – |
| 11 | GPT-6 Astra (high) API access | OpenAI | 92,1 % | – |
| 12 | GPT-6 Astra (medium) API access | OpenAI | 92,1 % | – |
| 13 | Claude Opus 5.5 (max) API access | Anthropic | 91,7 % | – |
| 14 | GPT-6.1 Sol (high) API access | OpenAI | 91,7 % | – |
| 15 | GPT-6.1 Sol (xhigh) API access | OpenAI | 91,7 % | – |
| 16 | Gemini 4 Argon (high) API access | Google DeepMind | 91,3 % | – |
| 17 | Claude Opus 5 (max) API access | Anthropic | 90,4 % | – |
| 18 | Claude Fable 5.2 (high) API access | Anthropic | 90,4 % | – |
| 19 | Claude Fable 5.1 (max) API access | Anthropic | 90 % | – |
| 20 | Claude Fable 5.1 (xhigh) API access | Anthropic | 90 % | – |
| 21 | GPT-5.6 Sol (xhigh) API access | OpenAI | 90 % | – |
| 22 | Claude Fable 5.2 (medium) API access | Anthropic | 89,6 % | – |
| 23 | GPT-6 Sol (max) API access | OpenAI | 89,6 % | – |
| 24 | Claude Fable 5 (max) API access | Anthropic | 89,2 % | – |
| 25 | Gemini 3.8 Flash (high) API access | Google DeepMind | 89,2 % | – |
| 26 | Claude Fable 5.1 (high) API access | Anthropic | 88,8 % | – |
| 27 | Claude Fable 5 (xhigh) API access | Anthropic | 88,3 % | – |
| 28 | Claude Opus 5 (high) API access | Anthropic | 88,3 % | – |
| 29 | Claude Fable 5 (high) API access | Anthropic | 87,5 % | – |
| 30 | Claude Opus 5.5 (medium) API access | Anthropic | 87,5 % | – |
| 31 | Gemini 4 Argon (medium) API access | Google DeepMind | 87,1 % | – |
| 32 | Claude Fable 5.2 (low) API access | Anthropic | 86,7 % | – |
| 33 | GPT-6.1 Sol (medium) API access | OpenAI | 86,7 % | – |
| 34 | Claude Fable 5.1 (medium) API access | Anthropic | 86,3 % | – |
| 35 | GPT-5.6 Sol (high) API access | OpenAI | 85,4 % | – |
| 36 | GPT-6 Astra (low) API access | OpenAI | 85,4 % | – |
| 37 | GPT-5.5 (xhigh) API access | OpenAI | 85 % | – |
| 38 | Gemini 3.7 Flash (high) API access | Google DeepMind | 84,6 % | – |
| 39 | Gemini 3 Deep Think Hosted access (no API) | Google DeepMind | 84,6 % | – |
| 40 | GPT-5.5 Pro (high) API access | OpenAI | 84,6 % | – |
| 41 | GPT-5.5 Pro (xhigh) API access | OpenAI | 84,2 % | – |
| 42 | GPT-5.6 Terra (max) API access | OpenAI | 83,9 % | – |
| 43 | Claude Sonnet 5.5 (xhigh) API access | Anthropic | 83,9 % | – |
| 44 | GPT-5.4 Pro (xhigh) API access | OpenAI | 83,3 % | – |
| 45 | GPT-5.5 (high) API access | OpenAI | 83,3 % | – |
| 46 | Gemini 3.8 Flash (medium) API access | Google DeepMind | 82,9 % | – |
| 47 | Claude Fable 5 (medium) API access | Anthropic | 82,5 % | – |
| 48 | Inkling 2 Small (xhigh) | Thinking Machines | 81,7 % | – |
| 49 | Claude Fable 5.1 (low) API access | Anthropic | 78,3 % | – |
| 50 | DeepSeek V4.1 Flash 1017 (max) Open weights (unrestricted) | DeepSeek | 78,1 % | – |
| 51 | GPT-6 Sol (xhigh) API access | OpenAI | 78,1 % | – |
| 52 | Gemini 3.8 Flash (low) API access | Google DeepMind | 77,5 % | – |
| 53 | Gemini 3.1 Pro Preview API access | Google DeepMind | 77,1 % | – |
| 54 | Claude Fable 5 (low) API access | Anthropic | 76,8 % | – |
| 55 | GPT-6.1 Sol (low) API access | OpenAI | 76,7 % | – |
| 56 | Claude Opus 4.7 (max) API access | Anthropic | 75,8 % | – |
| 57 | GPT-5.6 Terra (xhigh) API access | OpenAI | 74,2 % | – |
| 58 | GPT-5.4 (xhigh) API access | OpenAI | 74 % | – |
| 59 | DeepSeek V4.1 Flash 1017 (low) Open weights (unrestricted) | DeepSeek | 73,6 % | – |
| 60 | DeepSeek V4.1 Flash (max) Open weights (unrestricted) | DeepSeek | 72,9 % | – |
| 61 | Claude Opus 4.8 API access | Anthropic | 72,1 % | – |
| 62 | Gemini 3.5 Flash API access | Google DeepMind | 72,1 % | – |
| 63 | Claude Opus 4.8 API access | Anthropic | 71,7 % | – |
| 64 | GPT-5.5 (medium) API access | OpenAI | 70,4 % | – |
| 65 | Claude Opus 5.5 (low) API access | Anthropic | 70,1 % | – |
| 66 | Claude Opus 4.6 (120k thinking) API access | Anthropic | 69,2 % | – |
| 67 | GPT-6 Sol (high) API access | OpenAI | 68,9 % | – |
| 68 | DeepSeek V4.1 Flash 1017 (high) Open weights (unrestricted) | DeepSeek | 68,6 % | – |
| 69 | Claude Opus 4.7 (high) API access | Anthropic | 68,3 % | – |
| 70 | Claude Opus 4.7 (medium) API access | Anthropic | 67,5 % | – |
| 71 | DeepSeek V4.1 Flash (high) Open weights (unrestricted) | DeepSeek | 67,5 % | – |
| 72 | GPT-5.4 (high) API access | OpenAI | 67,5 % | – |
| 73 | Claude Sonnet 5.5 (high) API access | Anthropic | 67,4 % | – |
| 74 | Grok 4.6 (xhigh) API access | xAI | 67,1 % | – |
| 75 | GPT-5.6 Sol (medium) API access | OpenAI | 67,1 % | – |
| 76 | GPT-5.6 Terra (high) API access | OpenAI | 67,1 % | – |
| 77 | GLM-5.3-Flash (max) Open weights (unrestricted) | Z.ai (Zhipu AI) | 65,8 % | – |
| 78 | grok-4-20 API access | xAI | 65,1 % | – |
| 79 | Grok 4.6 (high) API access | xAI | 65,1 % | – |
| 80 | Gemini 3.7 Flash (medium) API access | Google DeepMind | 63,8 % | – |
| 81 | Claude Opus 4.8 API access | Anthropic | 62,2 % | – |
| 82 | Claude Opus 4.7 (low) API access | Anthropic | 62,1 % | – |
| 83 | DeepSeek V4 Flash 0731 (max) Open weights (unrestricted) | DeepSeek | 61,4 % | – |
| 84 | Grok 4.7 (xhigh) API access | xAI | 61,4 % | – |
| 85 | Claude Haiku 5.5 (xhigh) API access | Anthropic | 61,3 % | – |
| 86 | DeepSeek V4 Pro 0813 (max) Open weights (unrestricted) | DeepSeek | 61,3 % | – |
| 87 | Grok 4.6 (medium) API access | xAI | 61,3 % | – |
| 88 | DeepSeek V4.1 Flash (low) Open weights (unrestricted) | DeepSeek | 60,6 % | – |
| 89 | Claude Sonnet 4.6 (high) API access | Anthropic | 60,4 % | – |
| 90 | Gemini 3.6 Flash API access | Google DeepMind | 60,4 % | – |
| 91 | Kimi K3 (max) Open weights (non-commercial) | Moonshot | 60,4 % | – |
| 92 | DeepSeek V4 Pro 0813 (high) Open weights (unrestricted) | DeepSeek | 59,7 % | – |
| 93 | GPT-6 Astra (none) API access | OpenAI | 59,6 % | – |
| 94 | GPT-5.6 Luna (max) API access | OpenAI | 59,5 % | – |
| 95 | GPT-6 Luna (max) API access | OpenAI | 59,3 % | – |
| 96 | Grok 4.7 (medium) API access | xAI | 58,8 % | – |
| 97 | Grok 4.7 (high) API access | xAI | 58,3 % | – |
| 98 | Claude Sonnet 4.6 (max) API access | Anthropic | 58,3 % | – |
| 99 | GPT-6 Sol (medium) API access | OpenAI | 57,8 % | – |
| 100 | Muse Spark Open (xhigh) Open weights (restricted use) | Meta AI | 57,5 % | – |
