当前排名

推理

截至 2026-09-15 06:57:03 UTC，近 7 日推理体感排名：闭源组 — Claude Fable 5（86.5 分，n=66）；开源组 — GLM 5.3 Flash（79.3 分，n=38）。

每个场景展示 Top 3 具体版本，样本量满 5 条即可上榜；排名变化对比上一次发布的榜单。

闭源模型开源模型

#1 Claude Fable 5n=66 86.5▾

“Fable基本one shot everything”
来自知乎 · 2026-09-15

“These things you mentioned can be somehow avoided using less reasoning and more workflow (skills and etc) but that comes with a cost”
来自 Reddit · 英文 · 2026-09-14

查看模型详情 →

#2 GPT-6 Astran=156 79.7▾

“Astra is basically S+ tier and Fable is S tier; it can do stuff that Fable flubs but you really need a project that requires that much god-tier brainpower.”
来自 Reddit · 英文 · 2026-09-15

“pretty much not up for debate that Astra is the best model in the world”
来自 Reddit · 英文 · 2026-09-15

查看模型详情 →

#3 Claude Fable 5.1n=32 77.1▾

“推理不如Fable5.1”
来自知乎 · 2026-09-15

“you end up doing less rework since Fable agents are just much smarter”
来自 Hacker News · 英文 · 2026-09-14

查看模型详情 →

#1 GLM 5.3 Flashn=38 79.3▾

“it can't even consistently complete my annotation jobs without making up its own output schema or wrapping the json with code fences”
来自 Reddit · 英文 · 2026-09-15

“GLM 5.3 flash on the other hand is... Special...”
来自 Reddit · 英文 · 2026-09-14

查看模型详情 →

#2 Qwen3.8 27Bn=43 79.1▾

“qwen3.8 27b is pretty mind blowing to me. It can easily solve portswigger labs for example at q4_k_m.”
来自 Hacker News · 英文 · 2026-09-15

“Qwen3.8 27B for reasoning, coding, and agentic work”
来自 Reddit · 英文 · 2026-09-15

查看模型详情 →

#3 DeepSeek V4 Flashn=17 74.1▾

“DSV4's implementation was broken beyond 90k context”
来自 Reddit · 英文 · 2026-09-14

“DeepSeek Flash takes the spotlight”
来自 Reddit · 英文 · 2026-09-14

查看模型详情 →