Cursor released version 2.0 of its IDE on Wednesday, and with it Composer, the company’s first in-house coding model. Anysphere describes Composer as a frontier model roughly four times faster than similarly intelligent models, completing most turns in under 30 seconds, and trained with tools including codebase-wide semantic search. It was built with reinforcement learning and a mixture-of-experts architecture, and trained on interactive development challenges rather than static datasets.

Alongside the model, the interface is agent-first. Cursor 2.0 lets users run many agents in parallel without interference, using git worktrees or remote machines, and supports having several models attack one problem so the developer picks the best result. The company also begins addressing the bottlenecks it says emerge once agents write more code: reviewing it and testing it, the latter via a native browser tool the agent uses to check its own work.

The numbers come with a caveat: Cursor’s benchmark chart is internal. On it, Composer underperforms the “best frontier” on intelligence, outperforms top-tier open models, and far outpaces everything on speed. In Ars Technica’s reporting, a non-representative sample of developers said Composer is not ineffective but too expensive, given a perceived capability gap with the big models.

The same day, Cognition released SWE-1.5 for Windsurf, a frontier-size model it says runs at up to 950 tokens per second on Cerebras hardware — 6x faster than Haiku 4.5 and 13x faster than Sonnet 4.5. Both releases put speed at the center of the pitch.