Finding the Right Fit - TUA-Bench (OpenHands): leaderboard

TUA-Bench's 120 terminal-only computer-use tasks (office documents, email, web search, media and scientific workflows) run by the Finding the Right Fit harness study, every model inside the OpenHands agent harness at high reasoning effort with one counted run per task. Mean verifier reward in percent, higher is better (Li et al., 2026).

Metric: Task score (%), OpenHands harness, high reasoning effort, one run per task over 120 tasks. Source: huggingface.co. Saturation forecast: Around August 2028. 5 models tracked. Updated 5 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 5High65.23#26
2GLM-5.3High62.86#73
3GPT-6 AstraHigh61.01#10
4DeepSeek V4 ProHigh56.49#198
5Kimi K3High50.54#61

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

Interactive version: theaggregate.ai/benchmark?slug=finding-the-right-fit-tua-bench-openhands · How It Works · Data refreshed daily, snapshot 2026-11-20.