Bugbot, Cursor's automated code review agent, just got its biggest performance upgrade yet. The agent now runs in a median of 90 seconds, down from 5 minutes, costs 22% less per review, and surfaces 10% more bugs per run. For teams living inside GitHub pull request workflows, this is the kind of compounding improvement that quietly changes how fast code ships.
The numbers in full
The headline metrics are clean and measurable. Here is what changed across the three key dimensions:
| Metric | Before | After | Change |
|---|---|---|---|
| Average runtime | 5 min | 90 sec | 3x faster |
| Bugs found per run | 0.56 | 0.62 | +10% |
| Cost per run | $1.00-$1.50 | <$1.00 | -22% average |
In practice, 90% of Bugbot runs now finish in under three minutes. That is a meaningful shift for teams that were previously waiting on reviews before merging.
What is powering the gains
These performance gains are made possible by harness improvements and progress Cursor has made training Composer 2.5, which now powers Bugbot. Composer 2.5 is Cursor's in-house model, purpose-built for agentic coding tasks, and this is the first time it has been directly wired into the Bugbot review pipeline.
Bugbot uses a combination of frontier and in-house models to review code. The shift to Composer 2.5 as the primary driver is what unlocks both the speed and the improved detection rate simultaneously -- a combination that is harder to achieve than it sounds, since faster inference often comes at the cost of quality.
Don't miss what's next in AI
Join 300,000+ engineers and researchers who get the signal, not the noise.
- Full access to in-depth AI research breakdowns
- Be the first to know what's trending before it hits mainstream
- Daily curated papers, repos, and industry moves