Rankings
Snapshot 22 November 2026 · Scores copied from each benchmark's own leaderboard · How the CTO Index works
Score and price
| # | Model | Provider | ARC-AGI-2 | Price $/M |
|---|---|---|---|---|
| 1 | GPT-6.2 Astra | OpenAI | 95.8% | $8 / $40 |
| 2 | GPT-6 Astra | OpenAI | 95.0% | $10 / $50 |
| 3 | Gemini 4 Carbon | 94.2% | $2 / $10 | |
| 4 | GPT-6.1 Sol | OpenAI | 94.2% | $2 / $10 |
| 5 | Claude Fable 5.2 | Anthropic | 93.8% | $8 / $40 |
| 6 | Claude Opus 5.5 | Anthropic | 93.3% | $4 / $20 |
| 7 | Gemini 4 Hydrogen | 92.5% | $1.50 / $7.50 | |
| 8 | GPT-5.6 Sol | OpenAI | 92.5% | $4 / $20 |
| 9 | Gemini 4 Argon | 91.3% | $2 / $10 | |
| 10 | Claude Opus 5 | Anthropic | 90.4% | $5 / $25 |
| 11 | Claude Fable 5.1 | Anthropic | 90.0% | $10 / $50 |
| 12 | GPT-6 Sol | OpenAI | 89.6% | $2 / $10 |
| 13 | Claude Fable 5 | Anthropic | 89.2% | $10 / $50 |
| 14 | Gemini 3.8 Flash | 89.2% | $0.75 / $3.75 | |
| 15 | GPT-6.2 Sol | OpenAI | 85.8% | $2 / $10 |
| 16 | Gemini 3.7 Flash | 84.6% | $0.75 / $3.75 | |
| 17 | Claude Sonnet 5.5 | Anthropic | 83.9% | $2 / $10 |
| 18 | Inkling 2 Small | Thinking Machines | 81.7% | $0.30 / $1 |
| 19 | DeepSeek V4.1 Flash | DeepSeek | 78.1% | $0.15 / $0.60 |
| 20 | Claude Opus 4.8 | Anthropic | 72.1% | $5 / $25 |
| 21 | Grok 4.6 | xAI | 67.1% | $1.60 / $4.80 |
| 22 | GLM-5.3-Flash | Z.ai | 65.8% | $0.02 / $0.07 |
| 23 | Grok 4.7 | xAI | 61.4% | $2 / $6 |
| 24 | Claude Haiku 5.5 | Anthropic | 61.3% | $0.10 / $0.50 |
| 25 | Kimi K3 | Moonshot AI | 60.4% | $3 / $15 |
| 26 | GPT-6 Luna | OpenAI | 59.3% | $0.10 / $0.50 |
| 27 | Qwen3.8-27B | Alibaba | 42.4% | $0.40 / $3 |
| 28 | Beam 1.0 | Reflection AI | 23.5% | $0.30 / $1 |
Other benchmarks
Frequently asked questions
What does ARC-AGI-2 measure?
Novel visual puzzles that require inferring an abstract rule from a few examples. Each task is new, so memorised knowledge does not help.
Which model has the highest ARC-AGI-2 score?
GPT-6.2 Astra from OpenAI leads ARC-AGI-2 at 95.8%, ahead of GPT-6 Astra (95.0%).
Why does ARC-AGI-2 matter to engineering leaders?
A signal for how well a model handles problems it has not seen before, such as an unfamiliar codebase, an odd incident or a design trade-off with no template.
What are the limitations of ARC-AGI-2?
Top models now score above 90%, so the benchmark is near saturation at the frontier. Cost per task varies by more than 10x between models with similar scores.