Composer 2.5

Cursor's agentic coding model, announced 18 May 2026 as "our smartest, most capable coding model yet": steadier on long-running tasks and more reliable on complex instructions. Trained by Cursor via continued pretraining on Moonshot's open Kimi K2.5 checkpoint. Now the default pick; $0.50/1M input and $2.50/1M output, $3/$15 for the Fast variant.

Understands

text

This model cannot see images or video — you can only give it text. That matters if you planned to show it screenshots or documents with pictures.

Produces

text, code

Contents verified with the vendor: 2026-08-18

Included in subscriptions

No catalogue entry marks this model as part of a plan yet. It will appear here once the plan contents are confirmed with the vendor.