推理与知识
有可验证答案的难题，难在推理而不只是难在记忆。此处的分数在绝对值上偏低是刻意为之——选的就是这些还留有上升空间的基准。
按有多少个基准把该模型排进各自的前 10 名来排序，并列时看它的最佳名次。每一列保持各自的顺序，此处不做任何平均。
| 模型 | 进入前 10 | Artificial Analysis LLM Leaderboard GPQA Diamond | Artificial Analysis LLM Leaderboard Humanity's Last Exam | Artificial Analysis LLM Leaderboard CritPt |
|---|---|---|---|---|
 GPT-5.6 Sol openai | 3/3 | #3 94.1% | #3 49.5% | #1 32.3% |
 Claude Opus 5 anthropic | 3/3 | #5 93.7% | #2 54.9% | #4 29.1% |
 Kimi K3 kimi | 3/3 | #6 93.5% | #6 46.9% | #7 23.4% |
 Grok 4.6 xai | 2/3 | #1 94.9% | #8 44.1% | #12 19.7% |
| Claude Fable 5 (Opus 4.8 fallback) anthropic | 2/3 | #13 92.6% | #1 55.5% | #5 28.6% |
 Gemini 3.7 Flash google | 2/3 | #2 94.5% | #4 47.9% | #21 14.3% |
| GPT-5.6 Terra openai | 2/3 | #14 92.5% | #10 42.9% | #3 30.0% |
| Gemini 3.1 Pro Preview google | 2/3 | #4 94.1% | #5 47.0% | #15 17.7% |
 Qwen3.8 2.4T A95B alibaba | 2/3 | #7 93.5% | #13 42.4% | #10 20.0% |
| GPT-5.5 Pro openai | 1/3 | — | — | #2 30.6% |
| Gemini 3 Deep Think google | 1/3 | — | — | #6 25.7% |
 Muse Spark 1.2 meta | 1/3 | #23 90.4% | #7 45.5% | #16 17.7% |
| Grok 4.5 xai | 1/3 | #8 93.1% | #11 42.7% | #20 15.4% |
 GLM-5.2 zai | 1/3 | #27 89.5% | #17 41.1% | #8 20.9% |
 MiniMax-M3 minimax | 1/3 | #9 92.9% | #22 39.0% | #41 3.7% |
| Qwen3.8 Max alibaba | 1/3 | #12 92.7% | #9 43.0% | #11 20.0% |
| GPT-5.6 Luna openai | 1/3 | #19 91.1% | #21 39.5% | #9 20.6% |
 DeepSeek V4 Pro 0813 deepseek | 1/3 | #10 92.8% | #18 41.0% | #14 18.0% |
破折号表示该数据源未收录这个模型，而不是它得了零分。模型名录中没有条目的模型不在此处出现，因为它们无法跨数据源对齐——它们仍会显示在各自数据源的看板上，并带有标记。