OmniBehavior: leaderboard

OmniBehavior (2026) tests LLMs as simulators of real users: given the profile and long, real behaviour history of a user across the Kuaishou video, live-streaming, ads and e-commerce scenarios, the model predicts what the user does next. Overall score, the mean of the six behaviour metrics in the paper (binary-decision F1 in four scenarios, a watch-duration score and a judged dialogue score); higher is better.

Metric: Overall score, the paper's unweighted mean of its six behaviour metrics (four binary-behaviour F1 scores, a watch-duration score and a judged customer-service dialogue score) for predicting the real behaviour of 200 Kuaishou users (three months of interleaved video, live-stream, ads and e-commerce traces, 8,143 actions per user on average) from the user profile, history and current scenario, 6,000 balanced tasks, 32k context window, temperature 0.1; higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.544.55#160
2GLM-4.741.46#380
3Claude Sonnet 4.540.49#261
4GPT-5.239.07#185
5Kimi K2 090537.61#570
6OmniBehavior DeepSeek-V3 (checkpoint unspecified)37.42
7Claude Sonnet 436.87#410
8Claude Haiku 4.536.48#583
9GPT-4o36.27#751
10Gemini 3 Flash32.6#175
11Qwen 3 235B A22B 2507 Instruct32.11#616

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

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