OpenAI wants to know whether AI agents can do the unglamorous part of biology: reading a smudged gel, noticing a mislabeled control plate, or admitting that the data cannot support a conclusion. On November 6, 2026, the company introduced BioBasicsBench, a benchmark of 100 routine wet-lab tasks developed with Ligo, a biotech company focused on enzyme design. Its best model, GPT-6 Astra, earned 66.7% of the available credit at the highest common reasoning setting. OpenAI’s own write-up is just as interested in where Astra still fails.
Quick Take
- What: BioBasicsBench tests whether agents can interpret raw experimental data, catch errors and recognize insufficient evidence across 100 tasks from real Ligo experiments.
- Results: At xhigh effort, GPT-6 Astra scored 66.7% mean reward, GPT-6.1 Sol 62.0% and GPT-5.6 Sol 43.8% — with Astra using about half the output tokens of GPT-5.6 Sol.
- Caveats: Scores are partial rubric credit, not task success rates; only OpenAI models were tested; and the tasks have not been released.
- Why it matters: Lab AI has to get everyday evidence right before it can speed up discovery — and OpenAI says the benchmark will also inform biosafety and biosecurity tracking.
What happened
OpenAI announced the benchmark on its blog and on X, writing that “Accelerating biological research starts with getting the everyday work right.” Ligo CEO Ed Harris said the benchmark “reveals where models still struggle with routine experimental tasks.” According to OpenAI, existing benchmarks built around “charts and images selected for scientific papers may miss” the ambiguity of real bench work — failed experiments, ambiguous images and conflicting observations. BioBasicsBench is meant to measure exactly that.
Background: OpenAI’s push into science
The release fits a broader science push at OpenAI. R&D World notes that CEO Sam Altman said in a November 3 podcast that the company’s recent math results “really made me think that we are closer to this milestone than I expected us to be,” referring to a machine for scientific discovery. Days earlier, OpenAI president Greg Brockman forecast significant AI strides in biology and disease cures within 12 to 18 months. GPT-6 Astra, OpenAI’s top tier released in early September according to R&D World, has been at the center of several recent claims, including an MIT result on editing other models’ weights. BioBasicsBench is a more sober exercise: it measures routine competence rather than breakthroughs.
How BioBasicsBench works
The 100 tasks span five areas of routine wet-lab work: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. They range from reading the molecular weight of a DNA sample on a gel to detailed assay analysis. Ligo scientists wrote each task from their own experimental data and paired it with an expert-written rubric. Every task went through an internal review by a Ligo scientist who had not written it, an external review by a PhD-level wet-lab scientist, and a final internal review.
Inputs preserve the messiness of real lab evidence. By OpenAI’s count, 31 tasks include gels and blots, 27 protocols and notes, 20 tables and instrument exports, 18 bench and reagent photos, 14 microscopy data, 13 instrument screens and plots, 12 sequencing files, 7 protocol videos and 7 agar plates. Agents return a structured answer — a conclusion, measurement or recommendation, plus reasoning where required — and earn partial credit per rubric criterion, including for recognizing that “the evidence is insufficient.”
OpenAI evaluated seven of its own models across 37 model-and-effort settings, five attempts per task, for 18,500 results. The headline metric is mean reward: the average share of available credit, with 95% confidence intervals from resampling whole tasks.
| Model (xhigh effort) | Mean reward | 95% CI | Output tokens / task |
|---|---|---|---|
| GPT-6 Astra | 66.7% | 60.3–73.1% | 4,704 |
| GPT-6.1 Sol | 62.0% | 55.6–68.4% | 5,238 |
| GPT-5.6 Sol | 43.8% | 37.6–50.0% | 9,099 |
| GPT-6 Luna | 38.3% | 32.3–44.5% | 13,120 |
| GPT-5.5 | 35.4% | 29.7–41.0% | 11,135 |
| GPT-5.4 | 33.2% | 28.3–38.4% | 13,717 |
| GPT-5.2 | 31.2% | 26.2–36.3% | 27,620 |
What Astra gets right — and wrong
OpenAI credits GPT-6 Astra’s gains to “stronger native multimodal capabilities.” In one cloning task, GPT-5.6 Sol accepted an empty negative-control plate at face value; Astra noticed the rim label read “LB Kan50” — kanamycin, not the ampicillin used in the experiment — and rejected the control. In another, Astra recognized a microscope z-stack where GPT-5.6 Sol searched for a biological reason why cells seemed to vanish. Astra also refused to call cells dead from a dye signal when overlapping detection channels made brightness alone inconclusive.
The failure examples are equally instructive. In a colony PCR task, Astra grouped the gel lanes correctly but “assumes that a larger PCR product would identify successful cloning,” picking colonies 4, 6, 7 and 8 instead of the reference 5, 9 and 10 — which would lead to purifying the wrong plasmid. It generalized that vacuoles look darker while ignoring lighter ones in the same image. And it overcounted one colony plate by about 12%, which would inflate an estimate of transformation efficiency by the same proportion. OpenAI’s conclusion: “Higher benchmark scores do not yet mean models can make reliable judgments in the lab.”
Pros and cons
For users (scientists)
- Pro: A benchmark built on raw gels, plates and videos is far closer to daily lab work than paper-figure quizzes, and it rewards saying “not enough evidence.”
- Con: A two-thirds credit score still means frequent partial errors on routine tasks; scientists cannot yet hand off interpretation without checking.
For developers
- Pro: The results suggest multimodal perception, not longer reasoning, drives gains — Astra scored higher with fewer tokens than GPT-5.6 Sol.
- Con: The tasks are not public, so outside teams cannot reproduce results, test other models or build against the benchmark directly.
For companies (biotech and pharma)
- Pro: Rubric-based scores give lab leaders a measurable acceptance criterion before letting agents into analysis pipelines.
- Con: With only OpenAI models tested, procurement teams get no head-to-head comparison with Anthropic, Google or open-weight alternatives.
What X is saying
A new way to measure how AI agents handle the routine analyses of wet-lab science—BioBasicsBench. Developed with Ligo, a biotech company focused on enzyme design, the benchmark draws on 100 tasks from real experiments to test whether agents can interpret raw lab data, identify errors, and recognize when the evidence doesn’t support a conclusion.
— OpenAI (@OpenAI) November 6, 2026 · View on X
Excited to release BioBasicsBench, a benchmark we developed with @OpenAI to test whether AI agents can interpret real lab data, spot experimental errors, and recognize when the evidence doesn’t support a conclusion.
— Ed Harris (@edharris) November 6, 2026 · View on X
At the xhigh setting, Astra used 4,704 output tokens per task, about half the older model’s, so longer reasoning didn’t drive the gain.
— Rohan Paul (@rohanpaul_ai) November 6, 2026 · View on X
Several observers zoomed in on the scoring. “OpenAI’s top reported score is 66.7% with partial credit—not a task success rate,” Microsoft cloud and AI director Hamza Chishti wrote. Software engineer Omar liked that the benchmark tests for “there isn’t enough evidence” and added: “I’d love to see that ability broken out from the overall score.” The launch also spawned a viral hoax: a post with more than 750 likes claimed a browser-console command revealed a hidden “GPT-6.2 Astra” line at 89%. OpenAI’s post lists seven tested models, and no such model appears in its results. Analyst Michael Waitze noted the expectation gap more soberly: “We actually went into why today’s release was BioBasicsBench, not Astra.” For background on the next Astra version, see our report on GPT-6.2 Astra and Sol sightings.
What Grok says
Asked about the hidden-model claim, Grok called it “engagement bait,” describing a recipe in which a video injects “a fake window.__bioBasicsBench.reveal call” so that “a nonexistent GPT-6.2 Astra line at 89% pops up.” That is consistent with OpenAI’s published results, which include only seven models.
Critical analysis
BioBasicsBench gets the target right. Most science benchmarks test knowledge or paper comprehension; this one tests whether a model can be trusted with the boring evidence that decides whether an experiment worked. Rewarding “the data cannot support this” is especially valuable, because over-interpretation is a classic failure for both humans and models.
The limits are just as clear. OpenAI built, ran and reported the benchmark and evaluated only its own models; BenchLM notes OpenAI “has not released the tasks,” and keeps the scores out of its weighted rankings. One hundred tasks from one company’s lab is a narrow slice of biology, and rubric credit blurs the difference between a minor slip and a decision that wastes a week of bench time — such as selecting the wrong colonies. External replication and competitor results will determine whether this becomes a field standard or stays a vendor showcase. OpenAI also frames the benchmark as a way to track biosafety and biosecurity implications; it has not said how scores feed into its safeguards.
What’s still open
- Will OpenAI and Ligo release the tasks or an evaluation harness for outside models?
- How do Claude, Gemini and open-weight models score on the same tasks?
- How does performance break down for the “insufficient evidence” criterion specifically?
- How will BioBasicsBench results inform OpenAI’s biosafety thresholds?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on OpenAI’s November 6, 2026 BioBasicsBench announcement, R&D World’s coverage, BenchLM’s benchmark page, and posts on X from OpenAI, Ligo and independent observers. All scores are as reported by OpenAI and have not been independently reproduced. More on OpenAI.
i10x Perspective
The road to AI-accelerated biology runs through the lab bench, not the press release. BioBasicsBench is a useful honesty check: even OpenAI’s best model misreads controls and miscounts colonies. The next milestone is not a higher score from one vendor, but open tasks, rival models on the same leaderboard, and lab teams measuring agents against their own data before trusting them.
