According to DeepMind Beating, when retrieving genuine virus gene data, even cutting-edge AI models like GPT-5.5 or Claude often make mistakes. Due to the messy structure and incompatible interfaces of traditional biological databases (such as NCBI), large models struggle to consistently filter out qualified datasets. To make retrieval more reliable, Anthropic collaborated with the National Center for Biotechnology Information (NCBI) in the United States to develop the deterministic retrieval tool gget virus. Upon integration, the AI's data query accuracy directly increased to nearly 100%.
The research team tested various large models using their proprietary VirBench benchmark (comprising 40 pathogens and 120 retrieval tasks). Without tool assistance, the AI's retrieval accuracy fluctuated widely between 16.9% and 91.3%, and the results of running the same command multiple times were highly unstable. For example, in a query expected to return 266 sequences, Claude Sonnet 4 ran three times, retrieving only 106, 15, and 5 sequences respectively. In rigorous biological research, even a slight omission in genetic data can have disastrous consequences: reconstructing the Ebola virus's phylogenetic tree with incomplete data can erroneously push back the origin of the epidemic by nearly a century; when evaluating the effectiveness of antibody drugs like maftivimab and MBP134, AI may overlook key virus mutation sites, directly misleading treatment evaluations.
The newly developed gget virus is equivalent to building a "dedicated tunnel" under the complex biological database. Serving as a deterministic retrieval layer, gget virus automatically coordinates multiple APIs such as REST, Datasets, and E-utilities at the lower level, handling batch downloads, local filtering, and GenBank record extraction, and outputs standardized logs. Upon integration, the retrieval accuracy of all AI models surpasses 90%, with GPT-5.5 reaching 99.7%, completely eliminating inconsistencies from multiple runs. This also means that with the right tools, even low-cost models can provide high reliability. The research team believes that future biological databases should follow the development concept advocated by Andrej Karpathy—building for agents—to design for AI-centric scaled users and create a highly trusted biological context engine.
