LABBench2: How scientists test whether AI can do real biology research

LABBench2 is a test for AI systems that do biology research. It asks models to do tasks a lab scientist might do: find facts in papers, read figures, query databases and plan a cloning experiment. It was published in 2026 by Edison Scientific and FutureHouse researchers. It extends an earlier test called LAB-Bench.

It is the benchmark OpenAI named when it launched GPT-Rosalind, the model family behind our report on gpt-6-rosalind’s API price.

A quiz written for lab scientists, not for trivia fans

Think of a driving test. A written exam on road signs tells you something, but not whether someone can merge onto a highway. A general AI quiz is like the sign test. LABBench2 is closer to the road test: the questions come from the daily work of a biology lab.

One example from the original LAB-Bench paper, a database question: which gene is linked to distal renal tubular acidosis in the DisGeNet database but not in OMIM? The answer, WDR72, cannot be recalled from general knowledge. The model has to know how to look it up and compare two sources.

What the tasks actually ask the model to do

LABBench2 contains more than 1,900 tasks, according to its paper. Counts differ slightly between the paper and Edison’s announcement, so treat the total as approximate. The tasks continue the original themes:

• Literature retrieval: finding facts in published papers, often beyond the abstract.
• Figures and tables: reading charts and data tables from papers.
• Databases: pulling records from public biology databases.
• Protocol troubleshooting: spotting why a modified lab procedure would fail.
• Molecular biology: working with DNA sequences and designing cloning steps.

It also adds new areas, such as retrieving patents and clinical trial records and judging the quality of literature sources.

From multiple choice to open-ended work

The original LAB-Bench came from FutureHouse in 2024. Its authors, including Jon Laurent, Andrew White and Samuel Rodriques, built 2,457 multiple-choice questions in eight categories. Human PhD-level researchers answered a share of them as a baseline, with internet access and a DNA editor but a request not to use AI tools.

The paper reported that models often fell short of experts. It also flagged weaknesses. Weak wrong answers let a model succeed by elimination rather than reasoning, and the authors noted that on the hardest cloning questions, models seemed to do this. Open-response tests gave lower scores.

By 2026, the LABBench2 paper says, frontier models had improved enough that some LAB-Bench subcategories were near saturation or superhuman. The sequel changes the format:

• Many tasks use open-ended answers rather than multiple choice.
• Tasks often require tools, such as web search or code execution.
• Inputs arrive as text, uploaded files or material the model must retrieve itself.
• Figure and table questions are tested three ways: the image alone, the full paper, or retrieval of the source paper.

Its authors report that accuracy fell by roughly 26% to 46% compared with LAB-Bench, depending on model and subtask. The new test is much harder.

Where the benchmark shows up today

The paper reports that tools help most on retrieval-heavy tasks but do not solve them. Performance stayed weak on figure retrieval, supplementary information and database access. Models read figures and tables well when handed the image directly.

OpenAI used LABBench2 when it introduced GPT-Rosalind in April 2026. According to the-decoder, OpenAI said the model beats GPT-5.4 on 6 of 11 LABBench2 tasks. The biggest gain was on CloningQA, which requires fully designing DNA and enzyme reagents for a molecular cloning protocol. These are OpenAI’s own figures, and we have not verified them independently.

CloningQA is also a good picture of the test’s ambition. The LABBench2 paper says answers use a structured, machine-readable format that describes reagents and protocol steps, so a program can check the design. That is closer to a lab plan than a multiple-choice letter.

What a high score cannot tell you

A benchmark is a sample, not the whole job. Several limits apply.

• Coverage: The original paper states that LAB-Bench does not cover all of biology. LABBench2 is also a defined set of tasks.
• Small sections: The paper’s table lists only 14 base items for CloningQA, so a gain there rests on few questions.
• Self-reported results: Scores a lab reports for its own model are not the same as independent runs. The LABBench2 data is on Hugging Face and an evaluation harness is public on GitHub, which lets others check.
• Not a lab bench: A model that designs a cloning plan on paper has not run the experiment, and a plan can still fail in practice.
• Human comparison: In the original study, humans had tools that models lacked, and human coverage was about 69%, so gaps were hard to read.

What to watch next

Three things are worth following. First, whether independent groups publish LABBench2 results for models such as GPT-Rosalind, including those gated behind trusted-access programs. Second, whether the benchmark itself saturates, as its predecessor did. Third, whether benchmark gains turn into results in real labs, which no multiple-choice or machine-checked test can settle alone.

For readers following the pricing story, the benchmark explains what the premium would have to buy. The pricing page gives no benchmark data for gpt-6-rosalind, so any claim that it is worth the cost remains untested in public.