TUA-Bench: leaderboard
TUA-Bench (Chen et al., 2026) is 120 computer-use tasks an agent must complete entirely through a terminal: everyday work such as editing office documents, handling email, searching the live web and processing media, translated from OSWorld, plus scientific and engineering workflows built with domain experts. An in-environment verifier scores the final state, and the board reports the mean task reward in percent over five runs of every task, higher is better. Each entry is one agent scaffold, model and reasoning-effort configuration.
Metric: Success rate (%; mean of the verifier's scalar task reward over five runs of each of the 120 tasks; agent scaffold, model and reasoning effort as labelled). Source: arxiv.org. Saturation forecast: Around April 2027. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 60.1 | #48 (GPT-5.5) |
| 2 | Claude Opus 4.8 (Max) | 59.7 | #56 (Claude Opus 4.8) |
| 3 | Claude Opus 4.7 (Max) | 58 | #75 (Claude Opus 4.7) |
| 4 | GPT-5.5 (High) | 57.8 | #48 (GPT-5.5) |
| 5 | Claude Opus 4.7 (xHigh) | 55.4 | #75 (Claude Opus 4.7) |
| 6 | GPT-5.5 (Medium) | 51.5 | #48 (GPT-5.5) |
| 7 | Claude Opus 4.7 (High) | 49.9 | #75 (Claude Opus 4.7) |
| 8 | Claude Opus 4.7 (Non-reasoning) | 49.7 | #75 (Claude Opus 4.7) |
| 9 | Gemini 3.1 Pro (Preview) (High) | 49.3 | #91 (Gemini 3.1 Pro (Preview)) |
| 10 | Gemini 3.1 Pro (Preview) | 48.5 | #91 |
| 11 | Claude Opus 4.7 (Medium) | 45.7 | #75 (Claude Opus 4.7) |
| 12 | Claude Sonnet 4.6 (Max) | 42.8 | #158 (Claude Sonnet 4.6) |
| 13 | GPT-5.5 (Low) | 42.4 | #48 (GPT-5.5) |
| 14 | Claude Opus 4.7 (Low) | 41.2 | #75 (Claude Opus 4.7) |
| 15 | GPT-5.5 (Non-reasoning) | 36.5 | #48 (GPT-5.5) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tua-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.