Autonomous Science Bench^0.1

Evaluating AI on real workflows for autonomous scientific discovery

Run the benchmarkView the tasksContribute a task

Cite Autonomous Science Bench

If you find this work useful, please cite it.

BibTeX

@misc{autonomoussciencebench2026,
  title={Autonomous Science Bench: Evaluating AI for Autonomous Scientific Discovery},
  author={{Autonomous Science Bench Team}},
  year={2026},
  url={https://yibow.me/autonomous-science-bench},
}

Autonomous Science Bench 0.1 Leaderboard / 4 tasks

Tasks in biology, chemistry, materials and semiconductors · 1 trial each

Pilot results

Rank	Model	Agent	Resolution rate	Cost	Agent time
1	GPT-5.6 Terra(max)	Codex	75%3/4	$21.40	189 min
2	GLM-5.3 Flash	mini-SWE-agent	50%2/4	$6.59	369 min
3	Qwen3.8 27B	mini-SWE-agent	25%1/4	$9.76	365 min
3	Sonnet 5.5(high)	Claude Code	25%1/4	$2.54	148 min
5	GPT-5.6 Luna(max)	Codex	0%0/4	$3.07	251 min
5	Haiku 4.5	Claude Code	0%0/4	$1.75	24 min
5	MiniMax M3	mini-SWE-agent	0%0/4	$9.96	226 min

Resolution rate, share of trials that meet the task’s success criteria · Cost and agent time, totals per configuration. Task scores appear on each field’s tab. Ranked by resolution rate; ties share a rank.

Autonomous Science Bench0.1

Autonomous Science Bench 0.1 Leaderboard / 4 tasks

Autonomous Science Bench^0.1