Cursor's New Coding Model Just Changed Everything
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Cursor released Composer 2.5, a new coding model based on the Kimi K2.5 checkpoint, which the author discovered accidentally when Cursor auto-switched to it after exhausting premium model usage. The model performs comparably to frontier models like Claude Opus 4.7 and GPT-5.5 on coding benchmarks but costs roughly 14x less ($0.50 vs $7 per task). A live demo building a real-time collaborative whiteboard app showed Composer 2.5 completing the task in ~4 minutes with a working React/TypeScript app, while Opus took 15+ minutes and produced a broken result, and GPT-5.5 took ~10-15 minutes with a partially working app. The post also explains the concept of a 'coding harness' — the orchestration layer around an LLM that manages context, tools, and looping — arguing that Cursor's harness is superior to competitors like Claude Code, which amplifies model performance regardless of which underlying model is used.