Autonomous Science Bench^0.1

Evaluating AI on real workflows for autonomous scientific discovery

Run the benchmark Read the paper View the tasks Contribute a task

Cite Autonomous Science Bench

If you find this work useful, please cite it.

BibTeX

@article{wen2026autonomous,
  title={Autonomous Science Bench: Evaluating AI Agents for Autonomous Scientific Discovery},
  author={Yibo Wen and Han Liu},
  journal={arXiv preprint arXiv:2610.18449},
  year={2026},
  url={https://arxiv.org/abs/2610.18449}
}

Autonomous Science Bench 0.1 Leaderboard / 6 tasks

Pilot results

Rank	Model	Agent	Resolution rate	Cost	Agent time
1	Opus 5.5(max)	Claude Code	66.7%	$172.91	1142 min
2	GPT-5.6 Terra(max)	Codex	44.4%	$51.57	300 min
3	GPT-5.6 Luna(max)	Codex	27.8%	$6.91	377 min
4	GLM-5.3 Flash	mini-SWE-agent	11.1%	$25.68	1702 min
4	Sonnet 5.5(high)	Claude Code	11.1%	$7.84	163 min
6	MiniMax M3	mini-SWE-agent	5.6%	$28.37	747 min
6	Qwen3.8 27B	mini-SWE-agent	5.6%	$21.89	1118 min
8	Haiku 4.5	Claude Code	0.0%	$7.62	88 min

Resolution rate, share of trials that meet the task’s success criteria · Cost and agent time, totals per configuration. Each task scores its own metrics, listed on the Run page. Ranked by resolution rate; ties share a rank.

Autonomous Science Bench0.1

Autonomous Science Bench 0.1 Leaderboard / 6 tasks

Autonomous Science Bench^0.1