OpenAI Launches BioBasicsBench to Test AI on Wet-Lab Science

OpenAI launches BioBasicsBench for wet-lab AI testing

OpenAI just dropped a new benchmark called BioBasicsBench, built to stress-test how well AI agents handle the nitty-gritty, messy reality of wet-lab biology. Announced November 6, 2026, it packs one hundred tasks that force models to interpret raw experimental data, catch errors, and, crucially, recognize when the evidence just doesn't support a conclusion. They cooked this up with Ligo, a biotech outfit focused on enzyme design, aiming to make AI agents reliable enough to actually accelerate research at the bench.

Most existing benchmarks lean on polished charts and publication-ready images, which scrub out the ambiguity that defines daily lab life. Failed runs, fuzzy images, conflicting observations, that's the norm in a wet lab. An agent that can't navigate that noise risks steering a researcher toward wasted weeks and wrong conclusions. BioBasicsBench tries to capture that reality by pulling data straight from scientists' own experiments: gel and blot images, plate photos, microscopy stacks, sequencing reads, plate-reader exports, even videos of bench procedures. You also get reference materials like ladder charts and plasmid maps as context.

BioBasicsBench tasks and domains explained

The hundred tasks span five domains: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. They range from straightforward asks, like reading a molecular weight off a DNA gel, to detailed assay analyses requiring you to stitch together several evidence types. Every task was written by a scientist using their own data and paired with an expert grading rubric. Before inclusion, each passed three reviews: an internal check by a Ligo scientist who didn't author it, an external review by a PhD-level wet-lab researcher, and a final internal pass. The rubric breaks credit into individual criteria so the benchmark can tell a correct conclusion from a partially correct analysis.

GPT-6 Astra leads model evaluation results

OpenAI evaluated seven models: GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 Sol, GPT-6.1 Sol, GPT-6 Luna, and GPT-6 Astra. The run covered thirty-seven model and reasoning-effort combos, five attempts per task per setting, yielding eighteen thousand five hundred total results. Performance is reported as mean reward, average task score as a percentage of available credit, with ninety-five percent confidence intervals calculated by resampling whole tasks.

At the highest reasoning setting, labeled xhigh, GPT-6 Astra averaged sixty-six point seven percent (CI sixty point three to seventy-three point one). GPT-5.6 Sol, by contrast, landed at forty-three point eight percent (CI thirty-seven point six to fifty). The newer model hit that higher score while burning fewer output tokens per task, forty-seven hundred versus nine thousand ninety-nine for the older system. GPT-6.1 Sol split the difference at sixty-two percent. The rest clustered lower: GPT-6 Luna at thirty-eight point three percent, and the GPT-5 series models ranging from thirty-one to thirty-five percent.

OpenAI argues the results hint that stronger native multimodal capabilities are the key to cracking everyday wet-lab analysis. They flagged a specific failure mode where GPT-5.6 Sol overinterpreted a cell-viability assay. The model called cells dead based on brightness in a detection channel that also picked up signal from a second dye, meaning brightness alone couldn't justify the call. GPT-6 Astra caught the overlapping detection windows and the resulting ambiguity, then recommended better controls to separate the signals. That ability to discern the limits of what the data can actually show, rather than hallucinating conclusions, feels like a meaningful step forward.

Benchmark availability and implications for AI biology tools

For developers and researchers building AI tools for biology, the benchmark offers a concrete way to measure progress on the tasks that actually eat up a lab scientist's day. The inclusion of raw, uncurated data types, handwritten notebook pages, instrument screen photos, protocol videos, means models can't lean on clean formatting or idealized inputs. Rubric-based scoring also lets you see if a model is getting the right answer for the wrong reasons or missing critical nuances in its reasoning.

OpenAI hasn't released the benchmark itself for public download in this announcement, but the detailed breakdown of task construction, data types, and evaluation methodology gives the community a template for similar evaluations. Anyone integrating language models into lab workflows should treat the reported scores as a snapshot of current capability on a narrowly defined but representative slice of wet-lab work. The gap between best and worst remains wide, and even the top score leaves substantial room for improvement in image interpretation, control checking, and contextual reasoning.

If you're evaluating models for scientific use, consider running your own tests against the task categories described here, especially those combining multiple data modalities. Watch closely how models handle ambiguous or insufficient evidence; overconfidence in noisy data is a known risk in experimental settings. The benchmark also underscores that raw token counts and generic reasoning benchmarks may not predict performance on domain-specific multimodal tasks. As the field moves toward agents that can operate alongside scientists, evaluations grounded in real experimental messiness will matter more than leaderboard positions on sanitized datasets.

Frequently asked questions

What is BioBasicsBench and when was it announced?

BioBasicsBench is a new benchmark launched by OpenAI on November 6, 2026, designed to test AI agents on realistic wet-lab biology tasks. Developed with biotech company Ligo, it contains 100 tasks using raw experimental data like gel images, microscopy stacks, and sequencing reads to evaluate how well models interpret messy, real-world lab data.

How does BioBasicsBench differ from existing AI benchmarks for biology?

Unlike benchmarks using polished, publication-ready data, BioBasicsBench uses uncurated raw data from actual experiments, including failed runs, fuzzy images, and conflicting observations. Tasks span protein methods, imaging, quantification, nucleic acid methods, and molecular cloning, with expert rubrics that score partial credit for nuanced reasoning rather than just correct conclusions.

Which AI models were tested on BioBasicsBench and how did they perform?

OpenAI evaluated seven models: GPT-5.2, GPT-5.4, GPT-5.5, GPT-5.6 Sol, GPT-6.1 Sol, GPT-6 Luna, and GPT-6 Astra. At the highest reasoning setting, GPT-6 Astra scored 66.7% (CI 60.3-73.1), GPT-6.1 Sol scored 62%, GPT-5.6 Sol scored 43.8%, while GPT-6 Luna and GPT-5 series models ranged from 31-38%.

What key failure mode did the evaluation reveal in older models?

GPT-5.6 Sol overinterpreted a cell-viability assay by calling cells dead based on brightness in a detection channel that also captured signal from a second dye. It failed to recognize overlapping detection windows created ambiguity. GPT-6 Astra identified this limitation and recommended better controls to separate signals, demonstrating improved ability to discern data limits.

Is BioBasicsBench publicly available for download?

No, OpenAI has not released the benchmark itself for public download in this announcement. However, the detailed methodology for task construction, data types, and evaluation provides a template for similar evaluations. Researchers should treat reported scores as a snapshot of current capabilities on a representative slice of wet-lab work.

Source: OpenAI