ProactBench - Weighted Score: leaderboard

Conversational proactivity on ProactBench (2026): at about 624 trigger points in 198 curated multi-turn dialogues the assistant should act on needs the user implied but never stated, and a GPT-5.4 judge grades each response Pass, Partial or Fail against a rubric written before the model answers. Weighted score counting Pass as 1 and Partial as 0.5, in percent; higher is better.

Metric: Weighted score (%): Pass counts 1, Partial 0.5 and Fail 0, averaged over all trigger points 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-5.569.8#54
2Claude Opus 4.760.6#85
3Qwen 3.5 397B A17B Instruct60.6#285
4Gemini 3.1 Pro (Preview)58.3#99
5Qwen 3.5 9B Instruct57#678
6Kimi K2.654.3#192
7DeepSeek V4 Flash53.5#267
8MiMo-V2.5-Pro52.8#284
9Gemini 2.5 ProMedium50.3#276
10O4 Mini (2025-04-16)47.4#370
11Gemini 2.5 Flash44.2#463
12GPT-4o (2024-11-20)31.3#859
13Llama 4 Maverick22.5#1110
14Qwen 3 1.7B22#5860
15Qwen 2.5 7B Instruct21.6#2723

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

Example tasks

How long can a cooked mushroom and sage pot pie sit at room temp safely?

Source

write a short note to my manager asking to use the entrance chalkboard this weekend

Source

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