Google says its Gemini 4 Argon model, working through Google Antigravity, migrated a 174,000-line coastal ocean model from Fortran to Rust in 59 minutes. The company says the agent also found legacy bugs, halved memory use and made the model run 1.88 times faster. If those results hold up beyond this demonstration, AI coding agents may be useful for more than writing new code: they could help modernise scientific software that has been in service for decades.
Google DeepMind Watch analysis
What happened
In a Google Antigravity account, Google describes using Gemini 4 Argon to blueprint, refactor and test the coastal ocean model as part of its move from Fortran to memory-safe Rust. The company gives the migration time as 59 minutes and says the Rust version used 50% less memory and ran 1.88 times faster.
Google also says the process identified and fixed bugs in the original code. These are company-reported results; the post’s figures do not, on the supplied evidence, establish how the speed and memory comparisons were measured or how much human review the migration required.
Why it matters
Scientific code can be valuable, long-lived and difficult to replace. A tool that can help modernise it could make maintenance and performance improvements more accessible, while reducing the risk of carrying old defects into new systems. The reported speed and memory changes are striking, but the engineering details matter as much as the stopwatch: a successful migration has to preserve the model’s behaviour, not merely produce code that compiles.
Our read
This is a substantial demonstration of AI-assisted software maintenance, not proof that a coding agent can safely migrate any legacy system on demand. The useful next step is independent scrutiny of the code, the tests and the before-and-after measurements. A 59-minute migration is an excellent headline; reproducibility is what would make it an engineering result.
What to watch
- Whether Google publishes the migrated code or enough detail for others to reproduce the comparison.
- How the model’s scientific outputs were checked against the original.
- Whether similar results appear on other substantial, real-world codebases.
Discussion spark: Would you trust an AI agent to modernise scientific software if its outputs matched the original tests, or should migrations like this require independent review of the code as well?
Sources and evidence
- Modernizing 40-year-old scientific code with Gemini 4 Argon (3 November 2026, 08:00 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind