ProactBench - Critical: leaderboard

Pass rate on the Critical triggers of ProactBench (2026), turns 4 to 7 where the assistant must combine two or more disclosed details into a conclusion the user did not ask for; a GPT-5.4 judge grades each response Pass, Partial or Fail against a rubric written before the model answers. Percent; higher is better.

Metric: Pass rate (%) on the Critical triggers (turns 4-7: synthesising two or more disclosed anchors into a new conclusion) over the 198 curated dialogues of ProactBench (about 624 rubric-scored trigger points: 201 Emergent, 232 Critical, 191 Recovery), each trigger judged Pass, Partial or Fail by a GPT-5.4 judge against a rubric written before the model answers; dialogue histories are fixed from a curation run and the model regenerates only at the triggers, sampled once at temperature 0.7; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.575.4#54
2Gemini 3.1 Pro (Preview)65.5#99
3Kimi K2.664.4#192
4Qwen 3.5 397B A17B Instruct63#285
5Claude Opus 4.758.6#85
6MiMo-V2.5-Pro53.2#284
7Qwen 3.5 9B Instruct51.4#678
8DeepSeek V4 Flash50.6#267
9Gemini 2.5 ProMedium48.3#276
10O4 Mini (2025-04-16)44.8#370
11Gemini 2.5 Flash37.9#463
12GPT-4o (2024-11-20)18.5#859
13Llama 4 Maverick14.2#1110
14ProactBench Llama-3.2-8B (checkpoint unspecified)6.9
15Qwen 2.5 7B Instruct6.5#2723

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

Interactive version: theaggregate.ai/benchmark?slug=proactbench-critical · How It Works · Data refreshed daily, snapshot 2026-11-21.