OpenAI releases GPT-Rosalind, biology-focused LLM trained on 50 common research workflows

OpenAI has released GPT-Rosalind, a large language model trained specifically on biology research workflows, marking a departure from the generalized science models offered by other major AI companies.

Named after Rosalind Franklin, the model has been trained on 50 of the most common biological workflows and programmed to access major public biological databases, according to Yunyun Wang, OpenAI's Life Sciences Product Lead.

Core capabilities

The model is designed to address two specific research bottlenecks: the massive scale of genomic and protein biochemistry datasets accumulated over decades, and the highly specialized jargon across biology subfields that creates barriers for cross-disciplinary work.

According to OpenAI, GPT-Rosalind can connect genotype to phenotype through known pathways, infer structural or functional properties of proteins, suggest biological pathways, and prioritize potential drug targets. The company claims the model has been tuned to be more skeptical than typical LLMs, making it more likely to identify poor drug targets rather than defaulting to agreement.

OpenAI states the model demonstrates "reasoning" capabilities, defined as working through complex multi-step processes, and "expert-level" performance on unspecified biology benchmarks. Specific benchmark scores were not disclosed.

Access restrictions

Access is currently restricted to US-based entities through OpenAI's trusted access deployment program due to concerns about potential misuse, such as optimizing virus infectivity. A more limited Life Sciences Research Plugin will be made generally available, though timing was not specified.

Pricing details were not disclosed.

Unknowns

OpenAI did not address whether the model has solved the hallucination problem that affects other LLMs, particularly when explaining reasoning steps. The company also did not specify which biology benchmarks were used to evaluate "expert-level" performance, the model's parameter count, context window size, or training data cutoff date.

What this means

GPT-Rosalind represents a bet on vertical specialization over horizontal generalization in scientific AI models. While competitors like Google and Anthropic have released broad science-focused models, OpenAI's biology-specific approach could prove more effective for specialized research tasks—or it could sacrifice flexibility for marginal accuracy gains. The proof will come from real-world usage reports, which should emerge once researchers gain access through the restricted deployment program. The access limitations suggest OpenAI is proceeding cautiously with biosecurity concerns, though the specific threat model remains unclear given existing computational biology tools already available to researchers.