Lab agents can run the experiment. Novices still cannot follow them.

A benchmark for everyday lab work

OpenAI has introduced BioBasicsBench, a benchmark built with Ligo to measure how well AI agents handle routine wet-lab analysis. It covers 100 tasks across five areas: protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. The tasks use raw experimental data, and each was reviewed by an internal Ligo scientist, an external PhD-level scientist, and a final internal review. m:…0232

The headline result is a clear capability jump between model generations. GPT-6 Astra earned an average of 66.7% of available credit at the highest reasoning setting, against 43.8% for GPT-5.6 Sol. GPT-6.1 Sol reached 62.0%, while GPT-6 Luna scored 38.3% and GPT-5.5 scored 35.4%. OpenAI says the newer model handles photos, videos, and raw outputs from sequencing and plate-reader assays better, and applies tacit biological knowledge and experimental context more reliably. m:…0232

Even so, the benchmark is candid about what remains weak. Models still over-interpret data, ignore contradictory evidence, and fail to count colonies correctly in routine measurements. That matters because the point of the benchmark is practical: reliable interpretation of experimental data is what lets scientists move faster without drawing wrong conclusions from failed or ambiguous experiments. OpenAI also says it wants to use the benchmark to understand the implications of AI agents for biosafety and biosecurity. m:…0232

The novice gap is about bad advice

A separate reanalysis complicates a comforting story about why novices do not benefit more from AI help in the lab. The 2025 Active Site study involved 153 novice participants with minimal laboratory experience. The reanalysis argues that model error was a large part of the problem too: mid-2025 models frequently provided incorrect advice to participants in the study. m:…0484

Of 55 participants who received fragment selection advice from mid-2025 models, 37 received incorrect advice. Those models frequently gave wrong guidance on fragment selection, assembly methods, and laboratory procedures, and the errors directly contributed to task failure. Participants who received correct guidance on using Golden Gate assembly with the BsmBI enzyme were more likely to succeed in the hands-on portion of the task. m:…0484

The picture changes with newer models. Re-running the transcripts through 2026-era models showed a marked increase in accuracy, and GPT-5.6 Sol suggests correct fragments at a higher rate than the mid-2025 models. DeepSeek V4 Pro also performed substantially better than GPT-OSS, an open-weight model released in 2025. The authors say the 2025 study should not be treated as definitive negative evidence for AI uplift potential, and they warn that small changes in uplift could substantially change the risk landscape because the pool of potential novice actors is large. m:…0484

Models can build, but interpretation lags

A third strand of evidence comes from scientific modeling, where Anthropic researchers describe three steps: building, constraining, and interpreting. LLM-based agents can now write code for candidate models from natural-language descriptions, and simulation-based inference uses neural networks to reverse-engineer parameters from model outputs, which is computationally feasible for complex models where standard Bayesian inference is not. m:…2226

The argument is that fitting data is not the same as understanding it. A single best-fit model can hide other plausible parameter settings or model structures, so the researchers argue AI systems should provide a distribution of plausible models rather than a single fit. Interpretation is described as the most difficult step to automate, and the process of constructing a model provides more insight than the final model itself. m:…2226

The same caution shows up in a new benchmark for research-level problem solving. OpenFrontiers combines a corpus of 12,151 open research problems with a benchmark of 200 problems, of which 110 were resolved in 2026 and 90 remain open. Across seven LLM agents tested, closed-book rediscovery of solutions was rare, occurring in at most 7 of 40 resolved problems, and live-web success mostly reflected retrieval of already published solutions rather than genuine problem-solving. m:…7412

The safety check is uneven

The safety picture is similarly mixed. Benchling built BenchBench-Refusals to test how frontier models distinguish harmful from benign biology requests, using 160 prompts divided into benign, pseudo-hazardous, and hazardous risk levels. OpenAI models distinguish biology requests reliably, and refuse at roughly the same rates regardless of model size. m:…6917

Anthropic models are reliable if enrolled in the company's Life Sciences Verification Program, but otherwise refuse many routine requests. Its Fable 5.1 and Opus 5.5 refuse many routine requests without verification, while Sonnet 5.5 has looser restrictions even without it. Open-weight models rarely refuse any requests, including harmful ones. The study concludes that better models are not necessarily better calibrated for safety, so general capability is not a reliable indicator of safety calibration. m:…6917

Taken together, the five sources describe a field splitting in two directions. Capable models are getting measurably better at routine lab analysis, model building, and giving correct experimental advice. But the guardrails around that capability are uneven across providers, and the evidence that agents can genuinely rediscover open scientific results, rather than retrieve them, is still thin. For developers building lab tools, the implication is that the bottleneck is shifting from raw capability to verification: checking the model's reading of the data, its refusal behavior, and its interpretation before anyone acts on it. m:…0232 m:…7412 m:…6917