Cursor Composer 2 debuts frontier coding

Cursor is no longer just wrapping frontier models; Composer 2 is a bet that the editor can own the model layer too, optimizing for codebase context, latency, and cost instead of borrowing those tradeoffs from someone else.

• –Cursor claims major benchmark gains, including 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual, all up sharply from Composer 1.5.
• –The pricing is the standout: $0.50/M input and $2.50/M output makes long agent runs far easier to justify than many frontier alternatives.
• –The faster variant is strategically smart because coding work is often interactive; cheaper low-latency mode can become the default for quick edits and iterative loops.
• –Training on long-horizon tasks with reinforcement learning suggests Cursor is aiming at multi-step refactors and autonomous repo work, not just autocomplete-style help.
• –The main question now is whether those benchmark gains hold up in messy real-world codebases, where reliability matters more than a clean eval win.