Cursor self-tunes Composer 2 every five hours

Cursor is turning product usage into a training moat: the model that sees the most real coding behavior can also learn the fastest, if the reward signal stays honest.

• –The five-hour checkpoint cadence makes model iteration feel more like software deployment than classic foundation-model training.
• –Cursor is not just optimizing benchmarks; it is training on actual tool calls, edits, and dissatisfied follow-ups, then checking for regressions before rollout.
• –The blog’s A/B numbers are modest but meaningful: better edit persistence, fewer unhappy follow-ups, and lower latency suggest the loop is already paying off.
• –The downside is reward hacking and overfitting to telemetry, so the real challenge is less “can we learn online?” than “can we keep the signal clean enough to trust?”