We’re thrilled to introduce Composer 2.5, our smartest, most capable coding model yet, and it’s available today in Cursor!
Composer 2.5 is a substantial leap over Composer 2 in both raw intelligence and how it actually feels to work with. It’s noticeably better at sustained work on long-running tasks, follows complex instructions more reliably, and is genuinely more pleasant to collaborate with.
What’s new:
- Smarter on the hard stuff: We scaled training significantly and trained on 25× more synthetic RL tasks, with a focus on long-horizon coding work — the kind of multi-step problems that take hundreds of tool calls to get right.
- Better behavior, not just better benchmarks: Composer 2.5 introduces targeted RL with textual feedback that lets us shape specific behaviors like tool use, communication style, and effort calibration. These dimensions don’t always show up in benchmarks, but they’re the difference between a model that’s technically correct and one that’s actually pleasant to code with.
- Same base pricing as Composer 2: Composer 2.5 is priced at $0.50/M input and $2.50/M output tokens — identical to Composer 2. There’s also a faster variant with the same intelligence at $3.00/M input and $15.00/M output, which is less expensive than the fast tiers of other frontier models.
- 2x included usage for the first week — on us. Go wild.
- Built on Kimi K2.5: Like Composer 2, Composer 2.5 builds on Moonshot’s open-source Kimi K2.5 checkpoint, with our own continued pretraining.
It’s available now!
Composer 2.5 is the new default in the model picker (Composer 2 is still available if you want it). Just open the Cursor app or kick off a cloud agent.
If you want the deep technical story, including our new training methods and some genuinely fun examples of the model out-clevering our synthetic environments, the blog post is worth a read.
And this is just the start. Together with SpaceXAI, we’re training a much larger model from scratch with 10× more total compute on Colossus 2. We expect this to be a major leap in capability — stay tuned. 
We’d love for you to share your first impressions.