Capability, put to the test.
LiveBench scores objective tasks in reasoning, coding, mathematics, language and more. These results complement Arena human preference; the two measures are kept separate.
Benchmark edition 2026-06-25 · Checked 2026-10-14 10:18 UTC.
Source file updated 2026-10-13 16:42 UTC.
| Tested model / configuration | Overall / 100 |
|---|---|
| Claude Fable 5.1 Max Effortanthropic/claude-fable-5-1 | 83.41 |
| Claude 5.5 Opus Thinking Max Effortanthropic/claude-opus-5-5 | 83.22 |
| Claude Fable 5 Max EffortIdentity not mapped | 82.97 |
| GPT-6 Astra Max Effortopenai/gpt-6-astra | 82.16 |
| Claude 5.5 Opus Thinking xHigh Effortanthropic/claude-opus-5-5 | 82.06 |
| GPT-6.1 Sol Max Effortopenai/gpt-6.1-sol | 81.62 |
| Muse Spark 1.3 xHigh Effortmeta/muse-spark-1.3 | 81.59 |
| Muse Wonder 1.0 Max EffortIdentity not mapped | 81.16 |
| GPT-6.1 Sol xHigh Effortopenai/gpt-6.1-sol | 81.11 |
| DeepSeek V4.1 Flash Max Effortdeepseek/deepseek-v4.1-flash | 81.11 |
| GPT-5.6 Sol Max Effortopenai/gpt-5.6-sol | 81.05 |
| GPT-5.5 Thinking xHigh Effortopenai/gpt-5.5 | 80.19 |
| Claude 5 Opus Thinking Max EffortIdentity not mapped | 80.08 |
| Smaug AgenticIdentity not mapped | 79.51 |
| GPT-6 Sol Max Effortopenai/gpt-6-sol | 79.25 |
| Kimi K3moonshotai/kimi-k3 | 79.19 |
| Gemini 3.7 Flash HighIdentity not mapped | 78.83 |
| Qwen 3.8 Maxalibaba/qwen3.8-max | 78.46 |
| Grok 4.6 xHighIdentity not mapped | 78.04 |
| GPT-5.4 Thinking xHigh Effortopenai/gpt-5.4 | 77.97 |
| Muse Spark 1.2 xHigh EffortIdentity not mapped | 77.95 |
| GPT-5.6 Terra Max EffortIdentity not mapped | 77.94 |
| Claude Sonnet 5.5 xHigh Effortanthropic/claude-sonnet-5-5 | 77.75 |
| DeepSeek V4 Pro 0813Identity not mapped | 77.44 |
| Smaug FlashIdentity not mapped | 77.44 |
| Grok 4.7 xHighxai/grok-4.7 | 77.40 |
| Gemini 3.1 Pro Preview Highgoogle/gemini-3.1-pro-preview | 76.95 |
| Smaug MiniIdentity not mapped | 76.93 |
| DeepSeek V4 Flash Vision ExpIdentity not mapped | 76.76 |
| Claude 4.7 Opus Thinking xHigh Effortanthropic/claude-opus-4-7 | 76.53 |
| GLM-5.3zhipuai/glm-5.3 | 76.14 |
| Gemini 3.8 Flash Highgoogle/gemini-3.8-flash | 75.83 |
| Grok 4.5xai/grok-4.5 | 75.77 |
| Claude Sonnet 5.5 Max Effortanthropic/claude-sonnet-5-5 | 75.67 |
| Qwen3.8 27Balibaba/qwen3.8-27b | 75.27 |
| Claude 4.6 Opus Thinking High Effortanthropic/claude-opus-4-6 | 74.52 |
| Claude Haiku 5.5 xHigh Effortanthropic/claude-haiku-5-5 | 72.07 |
| GPT-6 Luna Max Effortopenai/gpt-6-luna | 72.03 |
| Mistral Large 4 Highmistral/mistral-large-4 | 71.84 |
| GLM-5.3 Flashzhipuai/glm-5.3-flash | 71.59 |
| deepseek-v4-prodeepseek/deepseek-v4-pro | 71.57 |
| Claude Haiku 5.5 Max Effortanthropic/claude-haiku-5-5 | 67.88 |
| Nemotron 3 Ultra 550B A55Bnvidia/nemotron-3-ultra-550b-a55b | 67.36 |
| Minimax M3minimax/MiniMax-M3 | 67.26 |
| Grok 4.3xai/grok-4.3 | 62.25 |
45 frontier and curated tested variants, from 70 upstream rows. Missing models are not assigned a score. Different effort settings remain separate.
Method & source
Overall is the equal mean of category averages; categories are the mean of their task scores. Incomplete results have no overall score. Benchmark editions can change the questions and difficulty, so scores across editions are not a continuous intelligence scale. First collection establishes when a result became known to this site; edition dates are not publication timestamps.
Official LiveBench results CSV ↗ · LiveBench leaderboard ↗ · All methods
Scores attributed to LiveBench. We publish a factual summary, not benchmark questions or model answers. An explicit leaderboard data licence has not been established.