← All benchmarks

BioResearchBench

Science

Office Hours Technologies' 50 PhD-written biology research tasks with realistic research data, packaged for Harbor. A task passes only if an answer meets every rubric criterion, and a model's score is the mean of its per-task pass fractions. Runs use Terminus 2 in Daytona, offline, with a 60-minute limit, at least 4 runs per task and a Claude Opus 5 grader, at high effort except Gemini 4 Argon at medium because of timeouts. The scores carry selection effects the publisher states: a task was accepted only if GPT-6 Astra failed it, tasks were revised or removed when newer models passed them, and timeouts are excluded from denominators, so a model that times out more is graded on fewer answers. The publisher also sells the benchmark commercially. Diagnostic for these reasons.

Benchmark source: Officehours

Top models (higher is better)