Google's ERA Writes Expert-Level Scientific Code — and Beat the CDC's COVID Forecaster

Google has detailed Empirical Research Assistance (ERA), an AI system that writes expert-level scientific software — and, in published benchmarks, outperforms the humans and institutions it was measured against. The work appears in Nature under the title "An AI system to help scientists write expert-level empirical software," released alongside Google's new Gemini for Science toolset.

How it works

ERA targets one of the most time-consuming parts of research: the iterative writing, testing and refinement of computational experiments. It pairs a large language model with tree search, systematically navigating the vast space of possible code solutions to maximize a defined quality metric. Rather than generating a single answer, it explores, evaluates and improves — closer to how a skilled research engineer actually works.

The results

The paper's benchmarks are striking:

• In bioinformatics, ERA independently discovered 40 new methods for single-cell data analysis, outperforming every human-submitted method on the public leaderboard.
• In epidemiology, ERA's forecasts outperformed the U.S. Centers for Disease Control and Prevention's own COVID-19 hospitalization forecasting ensemble.

These are not toy problems. They are the kind of applied, messy, domain-specific tasks where general-purpose models usually stumble — and where beating a purpose-built government forecasting model is a meaningful bar to clear.

Part of a broader push

ERA arrived the same day Google unveiled Gemini for Science, and it was one of two papers — alongside the Co-Scientist system — published in Nature simultaneously. Together they signal Google's intent to make AI a working member of the research team, not just a chatbot for literature review.

The release also intensifies a fast-moving contest: Anthropic recently launched Claude Science with Nobel laureate John Jumper, and Sakana AI's fully automated "AI Scientist" was itself published in Nature. The frontier labs are converging on the same wager — that the highest-value use of frontier models may not be writing emails or code for apps, but accelerating science itself. ERA's benchmark wins are early evidence the wager can pay off.