BioBasicsBench tests how AI agents interpret raw laboratory evidence. Its results bring attention to controls, context and uncertainty, and to the difference between earning partial credit on an analysis and providing dependable scientific assistance
A laboratory image may look straightforward to someone who knows what the experiment was meant to show. That expectation can be precisely what makes interpretation difficult. A faint signal may have several explanations, an apparently empty area may have been imaged poorly and a useful comparison may depend on a control recorded in another file. An AI assistant asked to interpret the material has to connect the observation with the experimental context. The important capability is to identify what the evidence supports and where another measurement or clarification is needed.
On 6 November 2026, OpenAI introduced BioBasicsBench with biotech company Ligo, an evaluation of 100 routine wet-lab tasks using raw experimental material. The inputs include images and instrument outputs, and the scoring uses expert-written criteria. OpenAI reports that GPT-6 Astra earned a mean 66.7% of available credit at its xhigh reasoning setting, compared with 43.8% for GPT-5.6 Sol. These are the developer’s evaluation results. The metric is partial credit averaged over tasks, so it cannot be interpreted as the share of laboratory experiments a system can complete successfully.
The choice of material makes the benchmark a useful subject for technical analysis. A figure prepared for publication is already the product of selection, processing and explanation. Raw evidence can leave the system with more work to do before it can make a defensible claim. It may need to read a label, establish which measurement belongs to which sample and decide whether two files describe the same condition. Testing those steps can reveal failures that a question about an established scientific fact would never expose.
From scientific knowledge to the work of interpretation
Earlier benchmark research provides context for this shift. The authors of LAB-Bench introduced more than 2,400 multiple-choice questions covering practical biology research capabilities, including literature use, figures, databases and biological sequences. That research broadened the evaluation of language models beyond textbook knowledge. Its question format also defines a particular testing environment. Choosing among supplied answers can exercise useful abilities while leaving open how a system will organise an analysis when the evidence is less curated and the acceptable answer has to be constructed.
The later LABBench2 paper describes nearly 1,900 tasks intended to present related capabilities in more realistic contexts, and supplies a dataset and evaluation harness. The authors report a substantial increase in difficulty relative to the earlier benchmark. That finding illustrates why performance on an established test can improve while a new test still exposes weaknesses. Each benchmark samples a different part of scientific work. A score becomes meaningful when the reader can identify the evidence supplied, the actions permitted and the form of answer that is rewarded.
An illustrative laboratory question shows the distinction. A system might know how a measurement is normally calculated and still be unable to decide which of several poorly labelled files contains the relevant values. Another system might extract those values correctly but apply the calculation to the wrong comparison group. A third might provide an acceptable number while overlooking a reason the comparison should not be made. Breaking these possibilities apart helps an evaluator diagnose whether the weakness lies in perception, context, calculation or judgement about the experiment.
Why the controls belong inside the reasoning
A control supplies a basis for interpreting a result. In a hypothetical imaging exercise, a bright region could reflect the target of interest, background signal or interference from another source. An assistant that reports the brightness faithfully may still make an unsupported biological claim. A useful analysis would connect the observation to the conditions that distinguish those possibilities. It would also make missing information visible. That response gives the researcher a reason to inspect or repeat a particular part of the work, rather than accepting an explanation simply because it matches the intended hypothesis.
This is where the shape of an answer matters. A researcher may ask for a number, a classification or a recommendation, but the system can need to state an assumption before supplying any of them. An illustrative response might identify a candidate measurement while marking it as provisional because a calibration record is unavailable. The evaluation has to decide whether recognising that limit counts as success. If a task always rewards a definite conclusion, it could encourage exactly the overinterpretation a scientist wants to avoid. Expert criteria can make that judgement explicit enough for others to examine.
There are technical implications for system design as well. A model may benefit from tools that inspect images, parse files or run calculations, but the additional steps create a record that needs to remain connected to the conclusion. A cropped image can make a small feature easier to see while removing a label. A numerical transformation can improve visibility while changing how intensities should be compared. The useful artefact is therefore an analysis whose evidence and transformations can be followed. That is a proposed design principle, with its effectiveness requiring evaluation on the intended tasks.
Read a benchmark score as a measurement with conditions
OpenAI’s methods account describes five attempts per task and reports confidence intervals calculated by resampling whole tasks. It also specifies reasoning-effort settings. These details help separate a result from a single lucky answer and make the comparison more interpretable. They leave other questions for deployment. Repeated attempts on the same task provide different evidence from performance on a new laboratory’s instruments, image conventions or experimental procedures. The benchmark is a sample of work, and a local evaluation would have to examine how far its results transfer.
Partial credit introduces a further distinction. A system could improve its score by making more of an analysis correct while still failing at the step that determines the practical decision. For an illustrative task, the assistant could identify the samples and calculate a ratio correctly but miss a control problem that undermines the interpretation. A higher mean score would still represent progress on parts of the work. A laboratory considering assistance would need to know the distribution of those failures and which errors a reviewer can catch reliably, alongside the overall average.
Comparisons also need a clear account of resources. More reasoning time, additional image inspections and extra tool calls can improve some answers while increasing delay and cost. A model that spends longer on a difficult task may be worthwhile if it prevents a costly repeat experiment. The same delay may be inconvenient for a time-sensitive observation. Evaluators could report latency, tool use and failure types together so that a researcher can judge the trade-off for a specific activity. Model names alone provide too little information to reproduce or interpret that choice.
What a laboratory could test next
A bounded pilot could begin with material whose interpretation has already been established by qualified researchers. The team could ask an assistant to produce a draft analysis, then compare it with the reference judgement and inspect disagreements. Cases could include imperfect images, missing context and situations where the proper answer is that the evidence is insufficient. These are proposed evaluation conditions, rather than a report of independent testing of BioBasicsBench. They would help identify whether the system is contributing to the actual task or merely recognising familiar presentations of it.
The next stage could test whether assistance changes the reviewer’s work. A correct draft may save time, but a plausible error can require a careful reconstruction of the analysis. The team would need to record both, including how often the reviewer accepts an incorrect result. An assessment of the combined process could distinguish errors the assistant makes from errors that survive review. That distinction matters because a system that performs well in isolation can still be difficult to supervise, especially if it presents uncertain conclusions with the same confidence as straightforward ones.
Public materials are another part of the evidence. The LABBench2 evaluation repository gives researchers an example of a benchmark with an inspectable implementation. For any comparison, access to task definitions, grading rules and execution conditions can help others find alternative explanations for a result. Some experimental material may require controlled access, but the limits should be clear. Independent reproduction of a reported score, a local test of usefulness and a study of long-term effects on scientific work would each answer a different question.
Making uncertainty useful
The possible contribution of AI assistance lies partly in making ordinary analysis more consistent and easier to review. A system could help locate an overlooked mismatch, organise evidence from several files or show which assumption a conclusion depends on. It could also introduce an error that a person would not otherwise have made. Those possibilities put the emphasis on the record the system leaves and the decisions it is permitted to influence. A pilot can establish a narrow useful role before a laboratory considers relying on it for more consequential judgements.
BioBasicsBench brings routine experimental interpretation into view as a technical capability worth measuring. The reported improvement is a reason to investigate further, with the developer’s scoring conditions kept attached. The strongest subsequent evidence would show how systems handle unfamiliar material, how they expose missing controls and how reliably researchers can verify their work. A laboratory assistant earns a place in that process when its output helps the team reach a more defensible conclusion, including the occasions when the available evidence calls for another question rather than an immediate answer.
Image credit: Petri dishes at a laboratory by Anna Loskutova, CC BY-SA 3.0. Illustrative photograph unrelated to BioBasicsBench.
