A new field report from OpenAI and academic partners shows that AI coding agents can modernize neglected scientific software with dramatic speedups — up to 60 times faster — while shifting the scientist's job from writing code to verifying that the science is still right.
The report documents eight case studies, mostly in biology, in which research groups used coding agents such as Codex, Claude Code, and GPT-5.x models. The projects ranged from basic maintenance to full rewrites in modern languages. RustQC delivered the largest gain, combining 15 separate quality-control tools into a single program and cutting runtime from 15 hours and 34 minutes to 14 minutes and 54 seconds — more than 60 times faster. HelixForge, a GPU-native replacement for a synthetic-genomics tool, ran the full pipeline 59.6 times faster than BamSurgeon, with its main compute step 98.6 times faster.
The most ambitious project rebuilt STAR, a 20,000-line C/C++ genomics aligner no longer actively maintained, from scratch in Rust. Tested on 10,000 yeast sequencing reads, the rewrite agreed with the original in 99.815% of single-end and 99.883% of paired-end cases. A rewrite of the R statistics package bayesm ran two to twenty times faster — but the first versions of two advanced methods contained subtle errors, including an inverted control parameter, that researchers only caught through calibration tests against thousands of synthetic datasets.
"With coding agents, it's quite easy to go fast; for now, to go far in science, there's still a need for expert guidance, understanding, taste, and care," wrote cyvcf2 developer Brent Pedersen. Philip Ewels, who led RustQC, described the agents as "eloquent, convincing, and confidently wrong in ways that are easy to miss." He never allowed the models to judge their own work, instead building an independent test harness. Across all projects, humans defined the goals, success criteria, and validation methods while agents handled implementation.
The report also sketches the economics: if agents resolved a quarter to half of installation problems across 100 packages, the saved research time would be worth $600,000 to nearly $5 million, and NumPy alone could save about 650 maintenance hours per year. But cheap rewrites create their own problems — fragmenting user communities and straining the limited time of expert maintainers. The report is part of OpenAI's broader push into science, led by a dedicated science team under Kevin Weil, who expects 2026 to be for science what 2025 was for software engineering.
