Odysseys - Hard: leaderboard

Odysseys (2026) benchmarks computer-use web agents on realistic long-horizon tasks composed from the browsing journeys of real users across several websites, graded item by item against per-task rubrics. This board is the perfect-rubric rate in percent on the hard tasks; higher is better.

Metric: Perfect-rubric rate (%: share of tasks with every rubric item satisfied) on the 109 hard tasks of the Odysseys long-horizon multi-site web tasks run in an OSWorld Ubuntu desktop with Google Chrome, at most 100 steps, maximum reasoning effort, each rubric item judged from the trajectory screenshots and actions by gemini-3.1-flash-lite-preview; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.6Max11#114
2Qwen 3.5 9B4.6#838
3Qwen 3.5 4B3.8#1270
4GPT-5.4xHigh3.7#138
5GPT-5.4 MinixHigh1.8#409
6Claude Sonnet 4.6Max1.8#179
7Qwen 3.5 35B A3B0#494
8uitars-1.5-7B0#7957

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=odysseys-hard · How It Works · Data refreshed daily, snapshot 2026-11-22.