Best AI models for reasoning
Reasoning tests ask a model to work through problems that cannot be answered by recall alone: graduate-level science questions, olympiad math, and expert exam questions with a single checkable answer. Across 16 models with at least one reasoning score, Claude Fable 5.1 has the highest index at 96.1. The strongest open-weight model is Kimi K3.1 Pro (88.7). Among models you can deploy on OpenLLM Buddy, Kimi K3 leads, ranked 8 overall.
Reasoning index
Average of each model's scores relative to the best published score on every test in this category (top = 100). Only tests a model reports are counted, and a model needs at least two of them to be ranked.
- OpenLLM Buddy catalog
- Open weights
- Closed API
- 1Claude Fable 5.1Anthropic · 2 of 4 tests96.1
- 2GPT-6 AstraOpenAI · 2 of 4 tests92.5
- 3Kimi K3.1 ProMoonshot AI · 2 of 4 tests88.7
- 4Gemini 3.8 FlashGoogle · 2 of 4 tests86.7
- 5DeepSeek V4 ProDeepSeek · 3 of 4 tests86.7
- 6Step 5 PreviewStepFun · 2 of 4 tests84.8
- 7Reflection Beam 1.0Reflection AI · 3 of 4 tests83.4
- 8Kimi K3Moonshot AI · 2 of 4 tests82.4
- 9Qwen3.6 27BAlibaba Qwen · 4 of 4 tests80.9
- 10GLM 5.3Z.ai (Zhipu) · 2 of 4 tests80.6
- 11DeepSeek V4.1 FlashDeepSeek · 2 of 4 tests75.9
- 12Qwen3.8 27BAlibaba Qwen · 2 of 4 tests70.3
- 13Gemma 4 31BGoogle · 3 of 4 tests69.8
- 14gpt-oss-120bOpenAI · 3 of 4 tests68.2
- 15Nemotron 3 UltraNVIDIA · 2 of 4 tests66.0
- 16Gemma 4 26B A4BGoogle · 3 of 4 tests63.2
10 more models report a single reasoning test and are listed in the table below without a rank: Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, Gemini 4 Argon, Grok 4.7, MiniMax M3.1 Flash, Qwen3.8 2.4T-A95B, Muse Spark Open, Tencent Hy4 Preview.
Raw scores by test
Published percentages. A dot means the model has no published score on that test.
| Model | GPQA | HLE | AIME | HMMT |
|---|---|---|---|---|
| Claude Fable 5.1 | 93.7 | 60.9 | · | · |
| GPT-6 Astra | 96.1 | 54.7 | · | · |
| Kimi K3.1 Pro | 94.2 | 51.1 | · | · |
| Gemini 3.8 Flash | 95.3 | 47.8 | · | · |
| DeepSeek V4 Pro | 90.1 | 42.7 | · | 95.2 |
| Step 5 Preview | 93.5 | 46.5 | · | · |
| Reflection Beam 1.0 | 90.6 | 36.0 | 97.7 | · |
| Kimi K3 | 93.5 | 43.5 | · | · |
| Qwen3.6 27B | 87.8 | 24.0 | 94.1 | 93.8 |
| GLM 5.3 | 91.7 | 42.3 | · | · |
| DeepSeek V4.1 Flash | 90.9 | 36.8 | · | · |
| Qwen3.8 27B | 89.2 | 30.8 | · | · |
| Gemma 4 31B | 84.3 | 19.5 | 89.2 | · |
| gpt-oss-120b | 80.1 | 14.9 | 95.8 | · |
| Nemotron 3 Ultra | 87.0 | 26.7 | · | · |
| Gemma 4 26B A4B | 82.3 | 8.7 | 88.3 | · |
| Claude Opus 5.5 | · | 64.4 | · | · |
| Claude Sonnet 5.5 | · | 56.9 | · | · |
| Claude Haiku 5.5 | · | 45.9 | · | · |
| GPT-6.1 Sol | · | 52.9 | · | · |
| Gemini 4 Argon | · | 57.1 | · | · |
| Grok 4.7 | · | 43.1 | · | · |
| MiniMax M3.1 Flash | · | 43.4 | · | · |
| Qwen3.8 2.4T-A95B | 93.5 | · | · | · |
| Muse Spark Open | 89.0 | · | · | · |
| Tencent Hy4 Preview | 92.3 | · | · | · |
Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works