Project record

CompBioBench v1

CompBioBench project timeline

How the project grew from initial Anycode runs into a reproducible comparison across models and execution environments.

25 completed replicates2,500 item traces9 milestones
2026

Initial Anycode evaluation

The first complete 100-item runs established a public baseline

CompBioBench began with three Codex GPT-5.5 Anycode runs. Their public leaderboard scores were 74%, 84%, and 83%, revealing substantial variation between repeated evaluations of the same model.

74%official84%official83%official
2026

Trace retention

Answers and execution traces became first-class project outputs

Each item was retained with its final answer and agent trace so disagreements could be reviewed rather than reduced to one aggregate score. Development submissions during this period reached 88%, 89%, and 91% on the public leaderboard.

88%official89%official91%official
2026

Experimental standardization

The evaluation protocol was made consistent across replicates

The project fixed the 100-item task set, execution conditions, reasoning settings, answer format, and retained artifacts. This created comparable formal replicates rather than a sequence of one-off runs.

2026

Anycode replicate matrix

Three model families completed three Anycode replicates each

Codex GPT-5.5, Codex GPT-5.6 Sol, and Codex GPT-5.6 Luna each completed three independent 100-item Anycode replicates under matched settings.

9completed replicates900item results
2026

Galaxy protocol

The same questions were evaluated through Galaxy

Galaxy runs used a fresh assigned history for every item. Agents could use standard Galaxy tools and create user-defined tools when a required operation was unavailable, while the scientific task and input data remained unchanged.

2026

DeepSeek-v4-pro-0813 expansion

Codex + DeepSeek-v4-pro-0813 completed both execution conditions

Three Anycode and three Galaxy replicates extended the comparison to a fourth model configuration, using high reasoning effort and the same 100-item benchmark.

6completed replicates600item results
2026

Verification and publication

Every published vector was checked item by item

Interrupted runs, answer files, traces, and Galaxy histories were reconciled before publication. Each published replicate contains 100 answers, and every answer links to a retained item trace.

2026

August release

Twenty-four completed replicates became available for review

The release included 2,400 item traces and 1,200 unlisted, link-accessible Galaxy histories, including Luna Galaxy results and token totals.

2026

Astra results

GPT-6 Astra scored 93% in Anycode

One Astra replicate adds 100 item answers and complete review directories. Sol Galaxy Replicate 3 scored 92% after an equivalent regulatory-element label was normalized for submission; the original agent trace is preserved.