Cursor Composer 1.5: 20x RL Scaling, Self-Summarization, 60% Latency Cut

Cursor Blog·high signalXBlueskyLinkedInCopy link

Cursor proprietary agentic coding model scales RL 20x beyond Composer 1. Post-training compute exceeds base pretraining. Adaptive thinking calibrated to difficulty. Self-summarization for long contexts. 60% latency reduction. 50% cheaper. Largest gains on hardest tasks.