推理与知识
有标准答案、难在推理而不在记忆的难题。这里的分数普遍偏低是正常的：专门选的是还有提升空间的基准。
按它在多少项基准里进入前 10 排序，一样多时看最好名次。每列保持各自的排名，不做平均。
| 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 6 Astra openai | 3/3 | #1 96.3% | #7 54.7% | #3 31.7% |
| GPT 5.6 Sol openai | 3/3 | #5 94.1% | #9 49.5% | #1 32.3% |
 Claude Fable 5.1 anthropic | 3/3 | #8 93.7% | #2 59.1% | #6 31.1% |
| Claude Opus 5.5 anthropic | 2/3 | — | #1 61.4% | #2 31.7% |
| GPT 6.1 Sol openai | 2/3 | — | #8 52.9% | #4 31.7% |
| Claude Sonnet 5.5 anthropic | 2/3 | — | #5 55.0% | #5 31.4% |
| Claude Opus 5 anthropic | 2/3 | #9 93.7% | #6 54.9% | #11 29.1% |
 Gemini 3.8 Flash google | 1/3 | #2 95.3% | #15 47.8% | #28 18.3% |
 Grok 4.6 xai | 1/3 | #3 94.9% | #22 44.1% | #25 19.7% |
| Gemini 4 Argon google | 1/3 | — | #3 57.1% | #14 27.1% |
| Gemini 3.7 Flash google | 1/3 | #4 94.5% | #14 47.9% | #40 14.3% |
| Claude Fable 5 anthropic | 1/3 | #17 92.6% | #4 55.5% | #12 28.6% |
| Gemini 3.1 Pro google | 1/3 | #6 94.1% | #16 47.0% | #31 17.7% |
 Muse Spark 1.3 meta | 1/3 | #7 94.1% | #11 48.7% | #16 26.0% |
| GPT 6 Sol openai | 1/3 | — | #13 47.9% | #7 30.9% |
| GPT 5.5 Pro openai | 1/3 | — | — | #8 30.6% |
| GPT 5.4 Pro openai | 1/3 | — | — | #9 30.0% |
| GPT 5.5 openai | 1/3 | #10 93.5% | #20 45.8% | #13 27.1% |
| GPT 5.6 Terra openai | 1/3 | #18 92.5% | #27 42.9% | #10 30.0% |
 MiMo V2.6 Pro xiaomi | 1/3 | — | #10 49.4% | #15 26.6% |
“—”表示该榜没收录这个模型，不是零分。还没识别出具体模型的条目不在这里出现，它们仍显示在各自的原始榜单上。