OpenAI just launched GPT-Rosalind, a large language model built specifically for biology and life sciences research. It is the company's first domain-specific model, it is restricted to US-based entities, and it is named after Rosalind Franklin, the chemist whose X-ray crystallography work was central to the discovery of DNA's double helix structure.

The model is designed to work through complex, multi-step biological reasoning: identifying drug targets, analyzing molecular interactions, and evaluating research hypotheses. OpenAI says it has been tuned to be skeptical rather than sycophantic, meaning it is more likely to tell a researcher their drug target is bad than to confirm whatever they want to hear.

That design choice is noteworthy. Sycophancy has been one of the most persistent and dangerous problems in AI-assisted research. Models that agree with whatever the user suggests are useless in scientific contexts where rigorous pushback is the entire point.

Restricted Access, Unclear Hallucination Track Record

Access is tightly controlled. Only US-based organizations can apply through OpenAI's trusted access deployment structure. The restriction is driven by biosecurity concerns: OpenAI is worried about the model being used to optimize viral infectivity or engineer biological agents. A more limited Life Sciences Research Plugin will be made generally available, but the full model stays behind gates.

What is unclear is whether OpenAI has solved the hallucination problem for scientific contexts. Ars Technica noted that the company has not addressed whether GPT-Rosalind can reliably explain its reasoning without fabricating intermediate steps. In biology, a hallucinated pathway does not just produce a wrong answer. It produces a wrong answer that looks indistinguishable from a real one until someone spends months and millions of dollars testing it in a lab.

The Real Competition: Anthropic Bought a Pharma Company

The timing is not accidental. Two weeks ago, Anthropic spent $400 million acquiring Coefficient Bio, a drug discovery startup. Anthropic put a pharma CEO on its board. It is building vertical integration into life sciences from the ground up: own the model, own the data, own the drug pipeline.

OpenAI is taking the opposite approach. Instead of buying pharma companies, it built a pharma model and is licensing access. The API play versus the acquisition play. Both are bets that AI-driven drug discovery is about to become one of the largest markets in technology. The question is which approach scales faster.

Other companies have released science-focused AI models, but they have been broader in scope. GPT-Rosalind is biology-specific, which could give it an edge in depth if not in breadth. Until independent benchmarks and real-world usage reports start arriving, the performance claims are marketing, not evidence.

The pharma AI race is now a two-horse competition between the company that builds the best model and the company that owns the pipeline. OpenAI chose the model. Anthropic chose the pipeline. Both cannot be right.

Reported by Reuters and Ars Technica. Additional analysis from Endpoints News.