Cursor released Composer 2.5 on May 18, 2026, a coding model built on Moonshot AI's Kimi K2.5 checkpoint that scores near Claude Opus 4.7 and GPT-5.5 on several benchmarks while offering lower token pricing.

  • The model uses the same foundation as Composer 2 but applies continued pretraining and reinforcement learning, with roughly 85% of final compute coming from Cursor's work.
  • Training introduced targeted RL with textual feedback to correct specific mistakes in long sessions, alongside 25x more synthetic tasks than its predecessor.
  • Composer 2.5 scores near Claude Opus 4.7 and GPT-5.5 on coding benchmarks, though it is restricted to the Cursor IDE with no public API.

The release highlights a shift toward harder reinforcement learning environments and cost-effective alternatives to frontier models for software engineering tasks.