LiveBench reasoning leaderboard
- Gemini 4 Argon leads
- 95.1%
- models measured
- 43
- lead over #2
- +2.4 pts
- median
- 87.8%
Leader on LiveBench reasoning · read October 21
On October 21, 2026, 43 of our catalog's models had a LiveBench reasoning result, and Google Gemini 4 Argon led with 95.1%, ahead of OpenAI GPT-6 Astra at 92.7%.
Short logic, spatial and theory-of-mind questions, each with a single correct answer graded against ground truth. Each score is the model's mean task score for the category, 0 to 100, from a single run of the current release, shown as a share.
Role in the index
LiveBench reasoning carries 20% of the Reasoning Index. A model missing it is scored on the other benchmarks of that index. Read the methodology
| # | Model | Result | Cost per task | Output tokens per task | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|---|---|
| 1 | Gemini 4 ArgonGoogle | 95.1% | $0.0781 | 7,571 | 81.0 | $10.00 | |
| 2 | GPT-6 AstraOpenAI | 92.7% | $0.3898 | 7,577 | 85.1 | $50.00 | |
| 3 | GPT-6.1 SolOpenAI | 92.6% | $0.0803 | 7,811 | 82.4 | $10.00 | |
| 4 | Claude Opus 5.5Anthropic | 92.2% | $0.2522 | 12,280 | 88.8 | $20.00 | |
| 5 | Claude Fable 5.1Anthropic | 91.7% | $0.4601 | 8,870 | 84.2 | $50.00 | |
| 6 | GPT-5.6 SolOpenAI | 91.7% | $0.1911 | 9,310 | 73.5 | $20.00 | |
| 7 | Claude Sonnet 5.5Anthropic | 91.6% | $0.2091 | 20,574 | 87.3 | $10.00 | |
| 8 | Claude Opus 5Anthropic | 91.2% | $0.1684 | 6,406 | 83.0 | $25.00 | |
| 9 | Grok 4.6xAI | 91.2% | $0.0873 | 14,139 | 65.7 | $6.00 | |
| 10 | Kimi K3Moonshot AI | 90.7% | $0.1951 | 12,770 | 62.3 | $15.00 | |
| 11 | GPT-5.6 TerraOpenAI | 90.6% | $0.1872 | 15,238 | 63.9 | $12.00 | |
| 12 | Muse Spark 1.2Meta | 90.0% | $0.0994 | 23,381 | 55.7 | $4.25 | |
| 13 | Claude Fable 5Anthropic | 89.7% | $0.6024 | 11,716 | 79.3 | $50.00 | |
| 14 | Muse Spark 1.3Meta | 89.7% | $0.1357 | 31,620 | 57.4 | $4.25 | |
| 15 | GPT-5.5OpenAI | 89.7% | $0.2892 | 9,457 | 71.3 | $30.00 | |
| 16 | Gemini 3.8 FlashGoogle | 89.3% | $0.1476 | 39,118 | 61.6 | $3.75 | |
| 17 | Claude Opus 4.8Anthropic | 89.2% | $0.6057 | 23,900 | 65.0 | $25.00 | |
| 18 | DeepSeek V4.1 Flash 1017DeepSeek | 89.1% | $0.0242 | 19,909 | 70.3 | $0.60 | |
| 19 | Claude Sonnet 5Anthropic | 88.7% | $0.2552 | 25,211 | 57.7 | $10.00 | |
| 20 | GPT-6 SolOpenAI | 88.7% | $0.1209 | 11,871 | 76.0 | $10.00 | |
| 21 | Qwen3.8 Max (0902)Qwen | 88.2% | $0.1315 | 21,515 | 62.4 | $6.00 | |
| 22 | Gemini 3.7 FlashGoogle | 87.8% | $0.0689 | 18,144 | 58.0 | $3.75 | |
| 23 | Qwen3.8 FlashQwen | 87.4% | $0.022 | 46,396 | Not benchmarked | $0.47 | |
| 24 | Grok 4.5xAI | 87.2% | $0.0877 | 14,539 | 58.6 | $6.00 | |
| 25 | DeepSeek V4.1 FlashDeepSeek | 86.7% | $0.0457 | 37,763 | 62.5 | $0.78 | |
| 26 | Claude Sonnet 4.6Anthropic | 86.4% | $0.313 | 20,620 | 57.3 | $15.00 | |
| 27 | GLM 5.3Z.ai | 85.8% | $0.2648 | 59,828 | 70.6 | $4.40 | |
| 28 | GPT-5.6 LunaOpenAI | 85.6% | $0.0277 | 22,753 | 58.0 | $1.20 | |
| 29 | DeepSeek V4 Flash Vision ExpDeepSeek | 85.4% | $0.0849 | 64,261 | Not benchmarked | $1.32 | |
| 30 | Mistral Large 4Mistral AI | 83.9% | $0.119 | 56,566 | 56.4 | $4.18 | |
| 31 | Step 5 PreviewStepFun | 83.8% | $0.0482 | 17,443 | 61.1 | $2.70 | |
| 32 | Grok 4.7xAI | 82.7% | $0.3851 | 63,769 | 66.7 | $6.00 | |
| 33 | Claude Haiku 5.5Anthropic | 82.4% | $0.0206 | 40,789 | 70.7 | $0.50 | |
| 34 | GPT-6 LunaOpenAI | 81.8% | $0.0214 | 42,629 | 57.3 | $0.50 | |
| 35 | Qwen3.8 27BQwen | 80.0% | $0.0749 | 24,786 | 74.2 | $2.55 | |
| 36 | MiMo-V2.6-ProXiaomi | 79.6% | $0.0237 | 26,730 | 65.4 | $0.87 | |
| 37 | Nemotron 3.5 Super VLNVIDIA | 78.8% | $0.0134 | 22,764 | Not benchmarked | $0.50 | |
| 38 | T InklingThinking Machines | 78.3% | $0.104 | 25,445 | 31.5 | $4.05 | |
| 39 | GLM 5.3 FlashZ.ai | 77.6% | $0.0138 | 27,228 | 57.7 | $0.50 | |
| 40 | MiniMax M3MiniMax | 74.5% | $0.0194 | 15,833 | 39.2 | $1.20 | |
| 41 | Muse Spark OpenMeta | 73.7% | $0.0427 | 9,861 | 47.4 | $4.25 | |
| 42 | Gemma 4 31BGoogle | 68.3% | $0.0038 | 9,019 | 39.0 | $0.40 | |
| 43 | Claude Haiku 4.5Anthropic | 61.7% | Not published | Not published | 35.6 | $5.00 |
Only models in our catalog are listed, and each appears once with its best published run. 43 rows were read from the source on October 21, 2026, and rows that match no catalog model are left out. Cost per task is what LiveBench reports the run spent on the category, divided by the number of tasks in it, in US dollars; LiveBench defines it the same way. Every other column is a figure the board publishes for the same run.
| Model | Result | Theory of mind | Zebra puzzles | Spatial | Logic with navigation |
|---|---|---|---|---|---|
| Google Gemini 4 Argon | 95.1% | 92.3% | 100.0% | 100.0% | 88.0% |
| OpenAI GPT-6 Astra | 92.7% | 84.6% | 100.0% | 98.0% | 88.0% |
| OpenAI GPT-6.1 Sol | 92.6% | 86.5% | 100.0% | 98.0% | 86.0% |
| Anthropic Claude Opus 5.5 | 92.2% | 84.6% | 100.0% | 98.0% | 86.0% |
| Anthropic Claude Fable 5.1 | 91.7% | 80.8% | 100.0% | 100.0% | 86.0% |
| OpenAI GPT-5.6 Sol | 91.7% | 84.6% | 100.0% | 100.0% | 82.0% |
| Anthropic Claude Sonnet 5.5 | 91.6% | 86.5% | 100.0% | 98.0% | 82.0% |
| Anthropic Claude Opus 5 | 91.2% | 78.8% | 100.0% | 100.0% | 86.0% |
| SpaceXAI Grok 4.6 | 91.2% | 90.4% | 96.5% | 98.0% | 80.0% |
| MoonshotAI Kimi K3 | 90.7% | 82.7% | 100.0% | 100.0% | 80.0% |
| OpenAI GPT-5.6 Terra | 90.6% | 86.5% | 100.0% | 100.0% | 76.0% |
| Meta Muse Spark 1.2 | 90.0% | 80.8% | 95.2% | 96.0% | 88.0% |
| Anthropic Claude Fable 5 | 89.7% | 84.6% | 100.0% | 96.0% | 78.0% |
| Meta Muse Spark 1.3 | 89.7% | 84.6% | 98.0% | 100.0% | 76.0% |
| OpenAI GPT-5.5 | 89.7% | 84.6% | 100.0% | 98.0% | 76.0% |
| Google Gemini 3.8 Flash | 89.3% | 76.9% | 98.2% | 100.0% | 82.0% |
| Anthropic Claude Opus 4.8 | 89.2% | 80.8% | 100.0% | 98.0% | 78.0% |
| DeepSeek V4.1 Flash 1017 | 89.1% | 88.5% | 100.0% | 98.0% | 70.0% |
| Anthropic Claude Sonnet 5 | 88.7% | 80.8% | 88.0% | 100.0% | 86.0% |
| OpenAI GPT-6 Sol | 88.7% | 84.6% | 100.0% | 96.0% | 74.0% |
| Qwen3.8 Max (0902) | 88.2% | 78.8% | 100.0% | 100.0% | 74.0% |
| Google Gemini 3.7 Flash | 87.8% | 82.7% | 94.5% | 100.0% | 74.0% |
| Qwen3.8 Flash | 87.4% | 80.8% | 98.8% | 100.0% | 70.0% |
| SpaceXAI Grok 4.5 | 87.2% | 82.7% | 94.0% | 100.0% | 72.0% |
| DeepSeek V4.1 Flash | 86.7% | 80.8% | 100.0% | 98.0% | 68.0% |
| Anthropic Claude Sonnet 4.6 | 86.4% | 75.0% | 94.5% | 100.0% | 76.0% |
| Z.ai GLM 5.3 | 85.8% | 88.5% | 88.8% | 96.0% | 70.0% |
| OpenAI GPT-5.6 Luna | 85.6% | 73.1% | 93.5% | 96.0% | 80.0% |
| DeepSeek V4 Flash Vision Exp | 85.4% | 78.8% | 98.8% | 96.0% | 68.0% |
| Mistral Large 4 | 83.9% | 80.8% | 93.0% | 94.0% | 68.0% |
| StepFun Step 5 Preview | 83.8% | 75.0% | 90.0% | 98.0% | 72.0% |
| SpaceXAI Grok 4.7 | 82.7% | 84.6% | 86.0% | 94.0% | 66.0% |
| Anthropic Claude Haiku 5.5 | 82.4% | 78.8% | 96.8% | 88.0% | 66.0% |
| OpenAI GPT-6 Luna | 81.8% | 73.1% | 96.0% | 96.0% | 62.0% |
| Qwen3.8 27B | 80.0% | 59.6% | 96.5% | 96.0% | 68.0% |
| Xiaomi MiMo-V2.6-Pro | 79.6% | 86.5% | 72.0% | 94.0% | 66.0% |
| NVIDIA Nemotron 3.5 Super VL | 78.8% | 76.9% | 80.2% | 98.0% | 60.0% |
| Thinking Machines Inkling | 78.3% | 65.4% | 88.0% | 94.0% | 66.0% |
| Z.ai GLM 5.3 Flash | 77.6% | 73.1% | 79.5% | 96.0% | 62.0% |
| MiniMax M3 | 74.5% | 76.9% | 61.0% | 94.0% | 66.0% |
| Meta Muse Spark Open | 73.7% | 76.9% | 40.0% | 100.0% | 78.0% |
| Google Gemma 4 31B | 68.3% | 73.1% | 36.0% | 96.0% | 68.0% |
| Anthropic Claude Haiku 4.5 | 61.7% | 44.2% | 58.5% | 90.0% | 54.0% |
The parts as LiveBench publishes them, for the same run as the result.