OpenAI BioBasicsBench: 100-task wet-lab agent benchmark with Ligo; GPT-6 Astra 66.7% vs GPT-5.6 Sol 43.8% at xhigh

Why it matters

【Key points】

On 2026-11-06 OpenAI released BioBasicsBench with enzyme-design firm Ligo: 100 routine wet-lab tasks that ask whether agents interpret raw experimental evidence, spot errors, and refuse unsupported conclusions—not paper-curated figures.

1. Domains: protein methods, imaging/cytometry, assays, nucleic acids, cloning. Inputs keep messy evidence (gels/blots, plates, z-stacks, sequencing, instrument screens, protocol videos). Scientists wrote tasks from their own data; Ligo internal + external PhD wet-lab + final review; rubric credit separates right conclusions from half-right analysis.
2. Official xhigh: GPT-6 Astra mean reward 66.7% (95% CI 60.3–73.1) vs GPT-5.6 Sol 43.8%; also GPT-6.1 Sol 62.0%, Luna 38.3% across seven models / 37 settings / five attempts (18, 500 runs). Astra better catches channel crosstalk, wrong-antibiotic “controls,” and z-stacks—but still fails colony-PCR logic, vacuole generalizations, and colony counts.
3. Motivation: paper-chart benchmarks miss failed runs and ambiguity; misreading controls wastes follow-ons and matters for biosafety.

【Our read】

1. The target moves from “read the paper figure” to “read the bench’s raw evidence,” opposite LifeVal on the same axis (everyday social tasks vs multimodal lab judgment). For lab workflows, that failure mode beats another curated biology QA set.
2. All numbers are OpenAI×Ligo self-tests; ~65% with lingering control errors is not a trusted assistant. Distinct from DeepMind cryptanalysis agents (e/1093) and OpenAI’s exploit pipeline (e/1089)—cyber/crypto vs biological evidence reading and refusal.
3. Watch open tasks, non-OpenAI scores, and biosafety uptake. Until then, OpenAI’s own wet-lab ruler—not an industry standard.

【Verifiable】

• Post
• X

Scores are OpenAI-reported. Related on-site but distinct: Google Research wet-lab grounding; OpenAI LifeVal (e/1082).