OpenAI introduces GPT-Rosalind: AI model specifically for biology research

OpenAI has announced a new AI model specifically for the life sciences. GPT-Rosalind – named after DNA researcher Rosalind Franklin – is tailored for biology, drug discovery, and translational medicine, which is the implementation of research findings in healthcare. The Frontier Reasoning model is intended to help researchers with evidence synthesis, hypothesis generation, and experiment planning, thereby accelerating early stages of drug development, which typically take ten to 15 years in the US.

As OpenAI explains in its blog, GPT-Rosalind understands in-depth connections in chemistry, protein engineering, and genomics. Additionally, the model can access more than 50 scientific tools and data sources via a new Life Sciences plug-in for Codex. This aligns with the strategy of serving other specialized domains with specific models, following GPT-5.4-Cyber for IT security.

Benchmarks show improvements

In internal tests, according to OpenAI, GPT-Rosalind surpasses its predecessors GPT-5, GPT-5.2, and GPT-5.4 in areas such as chemistry, protein biochemistry, phylogenetics, and tool usage. In the bioinformatics benchmark BixBench, it achieves a pass rate of 0.751 – the highest value among available models. On LABBench2, a benchmark for literature review and protocol design, it performs better than GPT-5.4 in six out of eleven tasks. In collaboration with Dyno Therapeutics, the model's top 10 predictions for RNA sequence prediction ranked in the 95th percentile compared to human experts.

OpenAI emphasizes that the model has been specifically tuned for skepticism to reduce hallucinations and overconfidence. Researchers are explicitly advised to treat the generated results as preliminary and validate them independently. Role-based access controls and a "Trusted Access Program" regulate who can use the model – currently, access is limited to qualified US enterprise customers.

Among the first partners are Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific. OpenAI announces GPT-Rosalind as the first model in a series: future versions are expected to better support long-term, tool-intensive research workflows. Whether the quantifiable benchmark improvements will actually translate into shorter drug development times remains to be seen in practice.

Competition in a growing market

GPT-Rosalind enters an increasingly competitive market. Anthropic offers a comparable product with Claude for Life Sciences, while Google DeepMind's AlphaFold specializes in protein folding. In German-speaking countries, the startup Puraite is working on explainable AI for evidence synthesis – a process that can manually take six months to two years. In contrast to these approaches, OpenAI positions itself as a provider of a broadly applicable model for entire research workflows, from sequence analysis to target prioritization.

However, GPT-Rosalind remains a closed-access model: OpenAI does not disclose weights, detailed error analyses, or internal reasoning steps. While researchers can see which external sources are included in the results through the integrated databases and tools, the model itself remains closed.

(vza)