DSAgentBench (Screenshot + A11y Tree): leaderboard
DSAgentBench (2026) evaluates end-to-end data-science workflows in a real OSWorld desktop across 275 tasks. Artifact checks verify execution and required outputs; about 10% of tasks also use visual judging after those checks. GPT-4o judges outputs from other models and Gemini 2.5 Pro judges GPT-4o outputs. Task success rate under screenshot plus accessibility-tree observation, in percent; higher is better.
Metric: Task success rate (%; percentage of 275 end-to-end data-science tasks (data acquisition, exploration, feature engineering, modeling, visualization, evaluation) whose final artifacts pass a deterministic Python evaluator (score >= 0.95), run in a real OSWorld desktop with VS Code, Jupyter and a terminal; temperature 0.1, 15-step budget; screenshot plus accessibility-tree observation). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 56.7 | #158 |
| 2 | GPT-5 | 29.81 | #163 |
| 3 | GPT-4o | 24.54 | #741 |
| 4 | Gemini 2.5 Pro | 20.81 | #262 |
| 5 | GPT-5 Mini | 19.03 | #349 |
| 6 | Claude Sonnet 4.5 | 9.21 | #244 |
| 7 | Claude Sonnet 4 | 4.64 | #397 |
| 8 | O4 Mini | 2.55 | #354 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dsagentbench-screenshot-plus-a11y-tree · How It Works · Data refreshed daily, snapshot 2026-11-12.