On November 6, 2026, OpenAI, together with enzyme design company Ligo, officially released BioBasicsBench—an AI evaluation benchmark covering 100 routine wet-lab tasks. Seven models were tested. GPT-6 Astra achieved a 66.7% average partial score under the highest reasoning setting (xhigh), the highest among all tested models, while the lowest-scoring GPT-5.2 scored only 31.2%, a gap of more than twofold. It is one of the largest publicly released biological science AI evaluations, and its source material is among the closest to real laboratory environments.
The core design orientation of this benchmark stands in sharp contrast to mainstream academic science benchmarks. Most existing biology evaluations rely on curated figures from papers, textbook illustrations, or multiple-choice questions with known answers, whereas BioBasicsBench uses raw evidence generated by laboratory scientists from their own experiments: gel and blot images, culture plate photographs, microscopy stack images, sequencing reads, plate reader export files, and video clips of actual bench work.
Five Domains, Three Review Rounds, and 19 “Cannot Determine” Questions
BioBasicsBench distributes the 100 tasks across five domains: protein methods (25 questions), imaging and cytometry (23 questions), quantification and detection (19 questions), nucleic acid methods (17 questions), and molecular cloning (16 questions). Each question was designed by Ligo scientists based on their own experimental data, accompanied by expert-written scoring rubrics, and passed through three rounds of review in sequence: an initial review by an internal Ligo scientist who did not participate in question creation, an external review by a practicing wet-lab PhD, and finally an internal final review. According to kingy.ai’s analysis, the model scores reported by OpenAI have not yet been independently reproduced by an external third party.
Scoring uses a partial-credit system—getting the conclusion right but the quantitative number wrong earns a different score from getting both the conclusion and the number wrong. In addition, 19 of the 100 questions specifically test whether models can recognize situations where “the evidence is insufficient to make a judgment.”
Results by Model Tier Across Seven Models
Under the xhigh reasoning setting, the seven models’ average partial scores were as follows (Source: OpenAI, November 6, 2026):
- GPT-6 Astra: 66.7% (95% confidence interval 60.3–73.1%), 4,704 tokens output per task
- GPT-6.1 Sol: 62.0% (55.6–68.4%), 5,238 tokens per task
- GPT-5.6 Sol: 43.8% (37.6–50.0%), 9,099 tokens per task
- GPT-6 Luna: 38.3% (32.3–44.5%), 13,120 tokens per task
- GPT-5.5: 35.4% (29.7–41.0%), 11,135 tokens per task
- GPT-5.4: 33.2% (28.3–38.4%), 13,717 tokens per task
- GPT-5.2: 31.2% (26.2–36.3%), 27,620 tokens per task
The improvement from 43.8% for GPT-5.6 Sol to 66.7% for GPT-6 Astra is about 23 percentage points, the largest single generational jump in the entire test sequence, while within the GPT-5 series, the improvement from GPT-5.2 to GPT-5.6 Sol was only about 12 percentage points. This suggests that the GPT-6 architecture provides a structural improvement in image reasoning capability, not merely linear scaling of parameter count.
GPT-6 Astra outputs 4,704 tokens per task, whereas GPT-5.4 and GPT-5.2 consume 13,717 and 27,620 tokens, respectively—GPT-5.2 uses nearly six times as many tokens as GPT-6 Astra—yet their scores are only 33.2% and 31.2%, respectively. According to cellcog.ai’s analysis, OpenAI attributes GPT-6 Astra’s token efficiency to its stronger native image-reading capability: the model can extract information directly from images without needing extensive textual reasoning to fill gaps in visual understanding.
Where GPT-6 Astra Got It Right—and Where It Went Wrong
cellcog.ai’s analysis presents specific cases disclosed in OpenAI’s release report. On the success side, GPT-6 Astra identified a control culture plate that had been labeled with the wrong antibiotic and caught an error in which Sol misidentified overlapping fluorescence channels as dead cells.
But in the same batch of tests, GPT-6 Astra selected the wrong colony in a cloning gel task and overcounted the colonies on one plate by about 12%.
OpenAI explicitly stated in its release report: “A higher benchmark score does not yet mean the model can make reliable judgments in the lab.”
Who Needs This Benchmark, and What Gap Does It Fill?
For procurement and R&D teams at life sciences companies, BioBasicsBench provides a reference framework that did not previously exist. Previously, AI vendors typically showed laboratory users general benchmarks of large model capabilities (such as the biology subset of MMLU and GPQA), but the evaluation scenarios in these benchmarks are quite far from real laboratory operations.
For AI R&D teams, this benchmark reveals a concrete technical direction: in wet-lab tasks, token efficiency and image reasoning capability are more predictive of actual performance than model parameter count. In the GPT-5 series, GPT-5.2 used nearly six times as many tokens as GPT-6 Astra but scored less than half as much.
For the competitive landscape, the release of this benchmark amounts to OpenAI setting a reference point for comparison in the life sciences AI race. Google DeepMind’s AlphaFold series is widely known for its leadership in protein structure prediction, but protein structure prediction is a relatively closed task with highly standardized input and output formats; BioBasicsBench, by contrast, measures the ability to interpret unstructured evidence in routine laboratory operations. There are currently no publicly available test results on BioBasicsBench from Google, Anthropic, or other major AI companies.
Actionable Conclusions for Developers and Enterprise Users
For teams planning to introduce AI agents into laboratory workflows, this benchmark offers several concrete starting points for action: First, BioBasicsBench includes 19 “insufficient evidence” identification questions, a capability that directly corresponds to the scenarios that most need to be guarded against in real deployments. In selection testing, prioritizing this type of task to assess candidate models is more valuable than total scores. Second, the score distribution across different task domains has not yet been fully disclosed (total scores and confidence intervals are known, but domain-level breakdowns have not been made public); in procurement decisions, buyers should ask vendors for subscores corresponding to their own business domains (such as imaging or nucleic acids) rather than looking only at average scores. Third, token consumption data is public in BioBasicsBench, and high token consumption often points to deficiencies in a model’s visual understanding.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接