BixBench3

BixBench3 evaluates the ability of AI agents to carry out analyses on the scale of complete research studies. Each task mirrors how scientists often use agents today: the scientist sets the research objective and a general methodological plan, then delegates implementation to the agent. Each task is based on a published computational biology paper. The agent must build a working analysis pipeline that analyzes the paper's raw data and produce structured artifacts supporting its findings. Those artifacts – such as protein abundance matrices or tables of differentially expressed genes – are graded programmatically against reference artifacts from the published study, measuring how well the agent executed the required analyses. We evaluated 13 frontier language models across 20 BixBench3 tasks encompassing the generation of 138 unique artifacts. Their average scores ranged from 0.00 for Gemini 3.1 Flash Lite to 0.48 for GPT 5.6 Sol. These were long and computationally intensive runs: an average attempt took 6.8 hours, processed 102 million tokens, and cost $43 while the largest consumed 1.07 billion tokens, ran for 24 hours, and cost $525.

Code and instructions on running the benchmark can be found in the BixBench3 repository.

The manuscript, Koch et al. 2026 "BixBench3 : Benchmarking AI agents on research-study-scale computational biology tasks" can be found here. Results of new models on BixBench3 will be updated on Edison Advances as models are released.

Task index

Task	Source Paper ID	Source Paper DOI	Artifacts
1	10.1101_2025.06.17.659900_v1	10.1101/2025.06.17.659900	12
2	10.64898_2026.02.11.704850_v1	10.64898/2026.02.11.704850	10
3	10.1101_2025.08.16.670679_v1	10.1101/2025.08.16.670679	4
4	10.64898_2026.02.03.703548_v1	10.64898/2026.02.03.703548	4
5	10.1101_2025.07.20.665670_v1	10.1101/2025.07.20.665670	10
6	10.1101_2025.07.08.663208_v1	10.1101/2025.07.08.663208	4
7	10.1101_2025.07.28.666515_v1	10.1101/2025.07.28.666515	4
8	10.1038_s42003-022-03654-9_v1	10.1038/s42003-022-03654-9	12
9	10_1186_s12915_024_01879_0_v1	10.1186/s12915-024-01879-0	14
10	10.1038_s41467-023-44243-6_v1	10.1038/s41467-023-44243-6	4
11	10.1016_j.crtox.2021.01.003_v1	10.1016/j.crtox.2021.01.003	5
12	10.1101_2025.06.02.657493_v1	10.1101/2025.06.02.657493	5
13	10.64898_2026.03.10.709901_v1	10.64898/2026.03.10.709901	4
14	10.1101_2025.07.21.665972_v1	10.1101/2025.07.21.665972	4
15	10.64898_2026.01.09.698641_v1	10.64898/2026.01.09.698641	7
16	10.1101_2025.07.31.667834_v1	10.1101/2025.07.31.667834	7
17	10.1101_2025.07.18.664654_v1	10.1101/2025.07.18.664654	4
18	10.64898_2026.01.02.697332_v1	10.64898/2026.01.02.697332	6
19	10.64898_2026.01.31.702960_v1	10.64898/2026.01.31.702960	14
20	10.64898_2026.02.04.703711_v1	10.64898/2026.02.04.703711	4