PIRA-Bench: leaderboard
Metric: Final score (%): average intent F1 on trajectories with actionable intents times the normalized false-positive score on pure-noise trajectories, on PIRA-Bench's 100 screenshot trajectories (mobile and desktop, each with noise frames and one of three user profiles); a Gemini-3-flash judge matches predicted intents to the human consensus intents; Naive setting (sliding window of 10 screenshots, prediction at the end); higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Seed 1.8 | 23.29 | #136 |
| 2 | Gemini 3.1 Pro (Preview) | 22.36 | #54 |
| 3 | Qwen 3.5 Plus | 21.18 | #123 |
| 4 | GPT-5.2 | 12.76 | #105 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=pira-bench · How It Works · Data refreshed daily, snapshot 2026-10-11.