This week: OpenAI and Ligo released BioBasicsBench, a 100-task report card for how well AI agents read raw wet-lab evidence — gels, colony plates, microscopy, and the messy bits papers leave out.

Elsewhere: How the bench is built, where models still trip, three open robotics and model drops this week, and a practical way to pressure-test your own lab assistant.

The New Prime Signal

Most AI biology demos look like a highlight reel: clean charts, labeled figures, an answer that fits the paper. Real benches look like a bad commute — smudged gels, a control plate with the wrong antibiotic on the rim, cells that vanish because someone scanned a z-stack instead of a time-lapse. On November 6, OpenAI introduced BioBasicsBench, built with Ligo, a biotech partner focused on enzyme design. The goal is blunt: measure whether agents can interpret raw experimental data, spot errors, and admit when the evidence cannot support a conclusion.

What shipped: BioBasicsBench has 100 tasks across five routine areas — protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Inputs keep the forms scientists actually see: agarose and SDS-PAGE gels, Western blots, plate photos, microscopy stacks, sequencing reads, plate-reader exports, and short videos of bench work. Agents return a structured answer; credit is scored against expert rubrics criterion by criterion, so a right conclusion with shaky reasoning still shows up. OpenAI reports evaluating seven of its own models across 37 model-and-effort settings, five attempts per task, for 18,500 graded results. At the shared xhigh reasoning setting, GPT-6 Astra earned a mean 66.7% of available credit (95% CI 60.3–73.1%), versus 43.8% for GPT-5.6 Sol (37.6–50.0%). Those are OpenAI's figures on OpenAI models — not an outside re-run.

Why it matters: Paper figures are curated. Bench work is not. Misreading a gel or trusting a broken control wastes the next experiment, not just the next chat reply. OpenAI's write-up argues existing benchmarks built from paper-selected charts miss the ambiguity and troubleshooting that fill a normal day. BioBasicsBench leans into that: overlapping dye channels, mislabeled plates, focus changes that reverse contrast. Astra's gains are framed as better multimodal reading and better restraint about what the data can show — still short of a reliable lab partner. If you care about AI for science beyond glamorous discovery demos, this is the unglamorous skill that has to improve first.

The fine print: Every score in the post is OpenAI grading its own models on tasks written from one partner lab's experiments. The public announcement does not link a downloadable task set, code release, or paper, so outside labs cannot yet rerun the bench or drop in other vendors' agents. Treat 66.7% as a vendor scorecard with a confidence interval, not a passport for unsupervised analysis. OpenAI itself shows failure modes that still hurt: assumptions overriding controls, preferred interpretations discounting contradictory cells, and colony counts off by roughly 12% on one worked example. Higher mean reward is progress; it is not permission to skip the human who knows which antibiotic was supposed to be on the plate.

We like report cards that grade the boring skills. Frontier models that ace a polished figure and then misread a rim label are like GPS that nails the highway and then parks you in a ditch. BioBasicsBench's honesty about z-stacks, overlapping channels, and overconfident colony counts is more useful than another victory lap. Watch for a public task drop; until then, use the posted examples as a checklist for your own assistant — and keep a scientist in the loop when the answer matters.

Raw evidence in, structured answer out — scored by rubrics that reward partial credit and honesty about insufficient data.

How it works: Ligo scientists wrote tasks from their own experimental data and paired each prompt with an expert grading rubric. Every task got an internal review by a Ligo scientist who did not write it, an external review by a PhD-level wet-lab scientist, and a final internal pass. Reviewers checked the prompt, evidence, rubric, author notes, and a graded model attempt. Most tasks mix skills: biological knowledge, image and table reading, error detection, and protocol sense. Data kinds overlap across tasks — OpenAI's table counts 31 gels-and-blots tasks, 27 protocol/notes/specs tasks, 20 table-and-export tasks, 18 bench photos, 14 microscopy, 13 instrument screens and plots, 12 sequencing files, plus colony plates, protocol videos, and supporting reference sheets such as ladder charts and plasmid maps. Mean reward is the average share of rubric credit across five attempts per task, with whole-task bootstrap confidence intervals.

Try it: Start with the OpenAI BioBasicsBench announcement and walk through the worked examples — overlapping fluorescence channels, the kanamycin rim label, the z-stack that looks like vanishing cells, and the colony-count mistakes. Those cases are a ready-made quiz for any multimodal assistant you already use: paste a messy gel or plate photo from a public methods figure and ask what the evidence can and cannot support. Secondary scoreboards such as BenchLM's BioBasicsBench page currently mirror the published aggregates as display-only because the task files are not public. If you need an open robotics stress test instead of a private bio quiz, jump to this week's AIRoA-0 release or AllenAI ToyBlocks — both ship weights or demos you can actually run.