PRL-Bench: leaderboard

PRL-Bench (2026) builds long-horizon theoretical and computational physics research tasks from 100 Physical Review Letters papers; a GPT-5 judge awards pre-set points for final answers and intermediate rubric checkpoints, mean of five runs, with a code interpreter and no search. This board is the score over all tasks on a 0-100 scale; higher is better.

Metric: Score (0-100) over all tasks, research tasks built from 100 Physical Review Letters papers; a GPT-5 judge awards pre-assigned points for final answers and intermediate rubric checkpoints, summed and normalized to 0-100, mean of five runs; code interpreter allowed, search disabled; higher is better. Source: arxiv.org. Saturation forecast: Around September 2027. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)44.27#102
2Qwen 3.5 Plus40.05#250
3Seed 2.0 Pro37.83#226
4Claude Opus 4.637.4#114
5GPT-5.437.38#138
6Kimi K2.533.89#272

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=prl-bench · How It Works · Data refreshed daily, snapshot 2026-11-22.