Cursor AI Updates: May 18, 2026

1. Cursor ships Composer 2.5, claiming Opus 4.7 and GPT-5.5 parity at $0.50/M input

Cursor. Cursor released Composer 2.5, a coding model built on Moonshot’s open-weight Kimi K2.5 checkpoint and trained on 25x more synthetic tasks than Composer 2, with roughly 85% of the compute budget going to additional pretraining and reinforcement learning. The company reports 79.8% on SWE-Bench Multilingual and 63.2% on CursorBench v3.1, levels it positions as matching Anthropic Opus 4.7 and OpenAI GPT-5.5 at $0.50/M input and $2.50/M output tokens, with a faster variant at $3.00/M input and $15.00/M output. The first week of usage is doubled as a launch bonus. Source

2. Composer 2.5 introduces “targeted RL with textual feedback” and a sharded Muon optimizer

Cursor. Composer 2.5’s writeup details two methodology shifts. The first is targeted reinforcement learning with textual feedback, which inserts localized hints at specific failure points inside long rollouts to fix the credit-assignment problem that plagues end-to-end RL on multi-step agent traces, with the hints aimed at tool-use and communication style. The second is on the training systems side: Cursor describes a “sharded Muon” optimizer using distributed orthogonalization plus a dual mesh HSDP layout that separates expert and non-expert weights, which it claims drives optimizer steps under 0.2 seconds on trillion-parameter mixtures. Cursor also disclosed a partnership with xAI to train a follow-up model on Colossus 2 using roughly 10x more compute. Source