Persona2Web (Pre-Execution): leaderboard
Metric: Success rate (times 100, 0-100): share of queries with full personalization and full intent-satisfaction scores; Persona2Web: 150 ambiguous queries for 50 synthetic users whose preferences appear only implicitly in long browsing and purchase histories (about 2,050 entries each), run on the real open web; the agent (AgentOccam or Browser Use, each extended with a planner, a history retriever and a generator that reasons over retrieved history) must resolve the missing website and preference from the history; a GPT-5-mini judge reads the full trajectory and reasoning against rubrics; pre-execution history access: the agent retrieves all relevant history before acting; higher is better. Source: arxiv.org. Saturation forecast: Around August 2028. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 13 | #240 |
| 2 | O3 | 10 | #121 |
| 3 | Gemini 2.5 Flash | 3 | #237 |
| 4 | Qwen 3 Next 80B A3B Instruct | 3 | #352 |
| 5 | Llama 3.3 70B Instruct | 2 | #520 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=persona2web-pre-execution · How It Works · Data refreshed daily, snapshot 2026-10-11.