Google’s Gemini 3 has set a new high-water mark on coding benchmarks, outperforming every publicly measured competitor on tasks that mirror day-to-day software engineering.
The gains are largest on multi-file reasoning, where the model has to hold an entire repository in context and make coordinated edits.
Why the benchmark matters
Synthetic puzzles have long overstated model ability. This suite instead draws from real pull requests, so the scores track more closely with what developers actually feel.
The jump on multi-file tasks is the part that surprised us most.
Google says the model is rolling out to its developer tools over the coming weeks.