ProactBench - Recovery: leaderboard

Pass rate on the Recovery triggers of ProactBench (2026), turns 8 to 10 after the user signals the task is done, where the assistant must add grounded forward-looking value tied to an earlier detail; 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 Recovery triggers (turns 8-10: grounded forward-looking value tied to an earlier detail after the user signals completion) 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 2028. 16 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.537.2#54
2Claude Opus 4.722.5#85
3DeepSeek V4 Flash18.3#267
4Qwen 3.5 397B A17B Instruct17#285
5O4 Mini (2025-04-16)16.2#370
6Qwen 3.5 9B Instruct15.4#678
7MiMo-V2.5-Pro13.1#284
8Gemini 3.1 Pro (Preview)11.5#99
9Gemini 2.5 ProMedium10.5#276
10GPT-4o (2024-11-20)8.9#859
11Kimi K2.67.4#192
12Qwen 2.5 7B Instruct6.8#2723
13Gemini 2.5 Flash5.2#463
14Qwen 3 1.7B4.2#5860
15Llama 4 Maverick2.1#1110

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

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