OpenAI has introduced BioBasicsBench, a new benchmark designed to evaluate whether artificial intelligence agents can interpret the raw experimental evidence generated in biological laboratories. Developed in collaboration with biotechnology company Ligo, the benchmark includes 100 routine wet-lab tasks involving DNA gels, microscope images, colony plates, sequencing files, laboratory notes and instrument readings.
The launch represents another step in OpenAI’s efforts to evaluate AI beyond conventional question-answering and software development. Instead of testing whether models can explain scientific concepts or summarise published research, BioBasicsBench examines whether they can interpret actual experimental evidence, identify mistakes and recognise when the available data does not justify a conclusion. GPT-6 Astra achieved an average score of 66.7% in the evaluation, demonstrating progress while leaving substantial room for improvement in scientific reasoning.
Key Takeaways
- OpenAI launched BioBasicsBench on November 6, 2026, in collaboration with biotechnology company Ligo.
- The benchmark contains 100 tasks covering routine wet-lab scientific work.
- It evaluates AI agents across five major areas, including protein methods, imaging, assays, nucleic acid methods and molecular cloning.
- GPT-6 Astra achieved a 66.7% average reward at the highest common reasoning-effort setting.
- GPT-6.1 Sol scored 62%, while GPT-5.6 Sol achieved 43.8%.
- OpenAI evaluated seven models across 37 configurations, producing 18,500 scored attempts.
- The benchmark tests whether AI can interpret raw evidence, recognise experimental errors and avoid unsupported conclusions.
- Remaining weaknesses include inaccurate colony counting, misinterpreting experimental controls and overlooking contradictory observations.
What Is OpenAI’s BioBasicsBench?
BioBasicsBench is an evaluation framework designed to measure how effectively AI agents handle everyday scientific analysis in wet laboratories.
Wet-lab research involves experiments conducted using biological materials, chemicals, cells, proteins and other physical samples. Scientists routinely collect evidence through laboratory instruments, imaging systems and experimental procedures.
Interpreting this evidence can be difficult.
A DNA gel image may contain unclear bands. A microscope recording may show cells moving through different focal planes. An experimental control may be incorrectly labelled, or a colony plate may contain overlapping bacterial colonies that are difficult to count accurately.
These situations require more than scientific knowledge. Researchers must understand the experimental procedure, recognise potential errors and decide whether the available evidence supports the expected conclusion.
OpenAI developed BioBasicsBench to test these capabilities.
The benchmark asks AI agents to analyse raw experimental information rather than relying only on clean charts and carefully selected images from scientific publications.
Its central objective is to determine whether AI can perform the analytical work that laboratory scientists regularly undertake.
What Does BioBasicsBench Test?
The benchmark contains 100 unique tasks across five areas of routine laboratory work.
| Research area | What AI agents must analyse |
|---|---|
| Protein methods | Protein gels, blots and related experimental evidence |
| Imaging and cytometry | Microscopy images, flow cytometry data and imaging results |
| Quantification and assays | Measurements, instrument exports and experimental calculations |
| Nucleic acid methods | DNA gels, sequencing files and molecular analysis |
| Molecular cloning | Plasmid maps, colony plates and cloning experiments |
The tasks can involve several different types of evidence at the same time.
For example, an AI agent may receive a microscope image, an experimental protocol and instrument measurements before being asked to interpret the results.
This requires the model to connect information from different sources rather than analysing each file independently.
GPT-6 Astra Scores 66.7% on BioBasicsBench
One of the most important findings from OpenAI’s evaluation is the performance improvement achieved by its newer models.
GPT-6 Astra recorded an average reward of 66.7% at the highest common reasoning-effort setting, outperforming the other tested models.
GPT-6.1 Sol followed with 62%, while GPT-5.6 Sol achieved 43.8%.
The results suggest that improvements in multimodal understanding and reasoning can help AI systems interpret complicated laboratory evidence.
AI Model Performance Comparison
| AI model | Average reward at xhigh |
|---|---|
| GPT-6 Astra | 66.7% |
| GPT-6.1 Sol | 62.0% |
| GPT-5.6 Sol | 43.8% |
| GPT-6 Luna | 38.3% |
| GPT-5.5 | 35.4% |
| GPT-5.4 | 33.2% |
| GPT-5.2 | 31.2% |
The scores represent mean reward, or the average percentage of available credit earned across the benchmark tasks.
They should not be interpreted as the percentage of experiments successfully completed or as a measure of how often a model produces a completely correct scientific conclusion.
OpenAI also reported 95% confidence intervals for the results, calculated by resampling complete tasks.
The evaluation covered seven models and 37 combinations of model and reasoning-effort settings. Each setting was tested five times on each of the 100 tasks, producing 18,500 scored attempts.
Why Is GPT-6 Astra’s Performance Important?
The difference between GPT-6 Astra and GPT-5.6 Sol is notable.
The newer model achieved 66.7% mean reward compared with 43.8% for the earlier model at the same reported reasoning-effort setting.
This indicates that newer AI systems are becoming better at combining visual information, scientific context and analytical reasoning.
However, a score of 66.7% also demonstrates that the model does not yet consistently handle every aspect of routine laboratory analysis.
Scientific work requires particularly high reliability because an incorrect interpretation can influence subsequent experiments and research decisions.
OpenAI Tests AI With Real Laboratory Evidence
A major difference between BioBasicsBench and many existing AI evaluations is the type of information provided to the models.
Traditional scientific benchmarks frequently use structured questions, clean datasets and images prepared for academic publications.
Real laboratory environments are less predictable.
Scientists work with photographs taken during experiments, handwritten notes, instrument displays, raw measurement files and imperfect images.
BioBasicsBench attempts to preserve this complexity.
Types of Data Included in the Benchmark
| Evidence type | Number of tasks containing this data |
|---|---|
| Gels and blots | 31 |
| Protocols, notes and specifications | 27 |
| Tables and instrument exports | 20 |
| Laboratory equipment and reagent photographs | 18 |
| Microscopy data | 14 |
| Instrument screens and plots | 13 |
| Sequencing files | 12 |
| Protocol videos | 7 |
| Agar plates and colonies | 7 |
| Reference sheets | 21 |
These categories overlap because individual tasks can contain several types of supporting information.
For example, a single cloning task may include a photograph of a laboratory plate, a DNA gel image, a plasmid map and an experimental protocol.
The model must interpret these materials together to reach a scientifically supported conclusion.
How Were the Tasks Created?
The tasks were developed using experimental data from Ligo’s own laboratory work.
Each task includes a question, supporting experimental evidence and expert-written grading criteria.
OpenAI says the tasks underwent several stages of review:
- An internal review by a Ligo scientist who did not create the task.
- An external review by a PhD-level scientist working in a wet laboratory.
- A final internal review before inclusion in the benchmark.
This process was intended to make the evaluation more realistic and ensure that the grading criteria reflected actual scientific requirements.
The benchmark also uses multiple criteria to award partial credit. A model can receive credit for a correct measurement or observation even if its final conclusion is incomplete.
AI Must Learn When Experimental Evidence Is Not Enough
One of the most important capabilities tested by BioBasicsBench is whether an AI agent can recognise the limitations of experimental evidence.
In scientific research, reaching the correct conclusion is not always about identifying the most likely explanation.
Researchers must also understand when an experiment has insufficient controls, when observations contradict each other or when a measurement cannot establish what it appears to show.
OpenAI identified this as an important area of improvement in GPT-6 Astra.
Recognising Ambiguous Microscope Images
One example involved microscopy images containing multiple fluorescent detection channels.
The channels had overlapping wavelength ranges, meaning that a signal detected in one channel could potentially be influenced by another fluorescent dye.
GPT-5.6 Sol interpreted a bright signal as evidence that certain cells were dead.
GPT-6 Astra recognised that the overlapping detection ranges created ambiguity and that brightness alone was not enough to establish the conclusion.
It identified the need for better experimental controls to distinguish between the signals.
This is important because scientific analysis requires understanding not only what an instrument measures but also what might interfere with that measurement.
Identifying an Incorrectly Labelled Laboratory Control
Another example involved a DNA cloning experiment.
A control plate was supposed to use a particular antibiotic, but its handwritten label indicated that a different antibiotic had been used.
GPT-5.6 Sol overlooked the incorrect label and based its conclusion on the apparent absence of colonies.
GPT-6 Astra identified the discrepancy and recognised that the control could not be treated as reliable evidence.
The example demonstrates why laboratory context matters.
An AI system may correctly analyse an image but still reach the wrong conclusion if it does not verify whether the experiment was conducted as intended.
Where AI Agents Still Make Scientific Mistakes
Despite the improvement in benchmark scores, OpenAI’s research also identifies several important weaknesses.
The examples demonstrate that even a model with stronger scientific reasoning can make mistakes when interpreting experimental evidence.
1. Incorrect Colony Counting
One task required GPT-6 Astra to count bacterial colonies on a laboratory plate.
The model struggled with closely spaced colonies, missed some visible colonies and incorrectly marked empty areas as colonies.
The combined result was an overcount of approximately 12% compared with the benchmark’s reference.
This type of error matters because colony counts can be used in calculations involving bacterial growth and transformation efficiency.
A seemingly small counting mistake can therefore influence the interpretation of an experiment.
2. Misinterpreting DNA Cloning Results
Another task involved interpreting colony PCR results.
The model identified two different band patterns but assigned the wrong experimental meaning to them.
It subsequently selected the wrong colonies for further work.
This example shows that recognising visual patterns is not sufficient.
The AI must also understand the controls and the biological context that determine what those patterns mean.
3. Ignoring Contradictory Observations
OpenAI also demonstrated a case in which GPT-6 Astra observed cells with both darker and lighter vacuoles in microscopy images.
However, the model gave greater weight to the darker examples and produced a general conclusion that overlooked the contradictory observations.
The researchers explained that changing the microscope’s focus could reverse the apparent contrast.
The problem was not simply image recognition. It involved deciding how much weight to give different observations and whether the evidence justified a general statement.
These examples show why AI-assisted scientific research requires careful verification.
Why BioBasicsBench Matters for the Future of AI in Biology
The launch comes as AI companies increasingly explore applications in scientific research, including drug discovery, protein engineering, genomics and medical research.
AI systems can already help researchers analyse large datasets, summarise scientific literature and generate potential research hypotheses.
However, biological research also depends on interpreting experimental results.
A model that understands scientific concepts but cannot reliably interpret the evidence collected in a laboratory has limited usefulness for many research workflows.
BioBasicsBench addresses this gap by evaluating the practical analytical capabilities needed to support scientists.
Potential Applications of AI in Laboratory Research
Faster experimental analysis: AI could help researchers interpret images, measurements and instrument outputs more quickly.
Error detection: Models could identify incorrectly labelled samples, inconsistent measurements and potential experimental problems.
Research troubleshooting: AI assistants could help scientists investigate why an experiment produced unexpected results.
Improved documentation: Models could organise laboratory notes, protocols and experimental observations.
Better research decisions: More reliable interpretation could help researchers decide which experiments deserve further investigation.
These applications could eventually reduce time spent on routine analytical work and allow scientists to concentrate on experimental design and more complex research questions.
However, BioBasicsBench does not demonstrate that AI agents can independently operate laboratories or conduct physical experiments. It tests their ability to analyse experimental evidence and provide conclusions.
OpenAI’s Broader Push Into Scientific AI Benchmarks
BioBasicsBench is part of a broader effort to evaluate AI capabilities in scientific research.
OpenAI has introduced several specialised benchmarks designed to test different aspects of scientific reasoning.
One example is LifeSciBench, which evaluates AI systems on expert-written and expert-reviewed life-science research problems.
Another is GeneBench-Pro, which tests how AI agents handle complex computational biology tasks involving ambiguous data, analytical decisions and scientific judgment.
These benchmarks address different research challenges.
BioBasicsBench focuses on routine wet-lab evidence, while GeneBench-Pro evaluates more complex computational biology analysis.
Together, they reflect a wider shift in AI research: moving beyond general knowledge tests towards specialised evaluations based on the actual work scientists perform.
The Bigger Picture
BioBasicsBench highlights an important transition in artificial intelligence. The next stage of scientific AI development will not depend only on how much information models can process or how well they explain scientific concepts. They will also need to interpret imperfect experimental evidence, understand laboratory procedures and recognise when a conclusion is not justified. OpenAI’s results show meaningful progress, particularly with GPT-6 Astra, but the remaining mistakes demonstrate that reliable scientific assistance is still a work in progress.
For the biotechnology industry, more capable AI systems could eventually accelerate routine analysis and help researchers identify experimental problems earlier. However, laboratory scientists will remain essential for validating findings, interpreting uncertainty and making consequential research decisions. The real measure of progress will be whether AI can consistently help scientists produce more reliable results, rather than simply provide faster answers.
Frequently Asked Questions
1. What is OpenAI’s BioBasicsBench?
BioBasicsBench is a benchmark developed by OpenAI and biotechnology company Ligo to evaluate how well AI agents interpret raw experimental evidence from routine biological laboratory work.
2. How many tasks are included in BioBasicsBench?
The benchmark includes 100 tasks covering five major areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods and molecular cloning.
3. Which AI model performed best on BioBasicsBench?
GPT-6 Astra achieved the highest average reward of 66.7% at the highest common reasoning-effort setting.
4. How did GPT-5.6 Sol perform?
GPT-5.6 Sol achieved an average reward of 43.8% at the same reported reasoning-effort setting.
5. What types of laboratory data does BioBasicsBench use?
It includes DNA gels, Western blots, microscope images, colony plates, sequencing files, instrument exports, laboratory photographs, videos, protocols and handwritten notes.
6. Can AI agents conduct laboratory experiments independently?
BioBasicsBench does not test independent physical laboratory operation. It evaluates whether AI agents can interpret experimental evidence and produce scientifically supported conclusions.
7. What are the main weaknesses identified in the study?
The research highlights problems involving colony counting, experimental controls, overlapping fluorescent signals and contradictory microscopy observations.
8. Why is BioBasicsBench important for biotechnology?
It provides a way to measure whether AI can assist with routine laboratory analysis, identify experimental errors and support researchers in making better-informed decisions.
Looking Ahead
OpenAI’s BioBasicsBench provides a new way to evaluate whether AI agents can become useful assistants in biological laboratories. Improvements in multimodal reasoning could help models interpret complex images, connect experimental observations and identify problems that might otherwise require additional human investigation. Further progress will depend on improving accuracy, contextual understanding and the ability to recognise uncertainty.
The long-term opportunity is to make scientific research faster and more efficient without compromising experimental reliability. AI may eventually support scientists across a wider range of laboratory workflows, from routine measurements to troubleshooting and data interpretation. However, achieving that goal will require models that can consistently distinguish reliable evidence from misleading observations and understand when human scientific judgment is still necessary.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
