Prim - Generation: leaderboard

Prim's direct-solving score: 182 expert-verified free-form mathematics problems from Humanity's Last Exam, answered zero-shot with chain-of-thought and scored by exact match on the final answer; higher is better (Prim, 2026).

Metric: Generation accuracy (%): exact match of the model's own zero-shot chain-of-thought final answer on 182 HLE-Verified mathematics problems; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 4 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4High67.03#138
2Qwen 3.6 27B52.75#352
3GPT-5.4 MiniHigh50.55#405
4Qwen 3.5 27B47.8#378
5GPT-5.4 NanoHigh44.51#651
6GPT-OSS-20BHigh43.41#1166
7Qwen 3.5 9B38.46#829
8Qwen 3.5 4B26.37#1235
9DeepSeek R1 Distill Qwen 32B18.13#1813
10DeepSeek R1 0528 Qwen3 8B17.03#2190
11DeepSeek R1 Distill Qwen 14B15.93#2706
12DeepSeek-R1-Distill-Qwen-7B13.19#7912

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=prim-generation · How It Works · Data refreshed daily, snapshot 2026-11-20.