If you thought AI coding assistants were already moving fast, buckle up. Cursor just dropped Composer 2.5, their smartest and most capable coding model yet.
While previous iterations were great at churning out boilerplate, Composer 2.5 represents a massive leap in handling long-horizon coding work. We are talking about the kind of complex, multi-step problems that take hundreds of tool calls to get right.
Here is everything you need to know about the update, the tech behind it, and why it is a big deal for developers. 👇
🧠 Smarter on the Hard Stuff
The biggest bottleneck with AI coding tools has always been sustained context. They start strong, but lose the plot after a few files.
Cursor tackled this head-on by scaling their training significantly. Composer 2.5 was trained on 25x more synthetic RL (Reinforcement Learning) tasks than its predecessor.
However, simply scaling data is not enough. When a model's rollout spans hundreds of thousands of tokens, it becomes incredibly difficult to assign credit—meaning the AI struggles to know which specific decision helped or hurt the outcome.
To fix this, Cursor introduced targeted textual feedback during RL.
Instead of waiting for the end of a rollout to penalize the model, they provide feedback directly at the exact point where the model messed up. For example, if the model makes a bad tool call or provides a confusing explanation, it receives a localized hint describing the desired improvement. This shapes crucial behaviors like communication style and effort calibration, making the AI genuinely more pleasant to collaborate with.
💻 The Code: How to Leverage Long-Horizon Agents
Because Composer 2.5 is built for sustained work, you can give it much more complex architecture tasks. Instead of asking for a single function, you can set up a .cursorrules file to define a long-running agentic workflow.
Here is an example of how you might instruct a long-horizon model like Composer 2.5 to autonomously refactor a legacy codebase:
# .cursorrules
You are an expert systems architect. Your task is to refactor the legacy `auth` module into a modern, scalable service.
## Workflow Execution Steps:
1. **Analyze:** Read all files in the `/src/legacy_auth` directory.
2. **Plan:** Draft a migration plan and wait for my approval before writing code.
3. **Execute:** Implement the new JWT-based auth flow across all middleware.
4. **Test:** Generate unit tests for the new implementation.
5. **Verify:** Run the tests using the terminal tool. If any fail, autonomously fix the errors until all tests pass.
*Note: If you encounter a missing dependency, use the terminal to install it.*
With Composer 2.5's improved tool use and behavioral shaping, it can actually execute a multi-step loop like this without hallucinating halfway through.
🏗️ The Elephant in the Room: The Kimi Base
There was a lot of community noise around Composer 2 being built on top of Moonshot AI's open-source Kimi K2.5 checkpoint. Cursor acknowledged this, and confirmed that Composer 2.5 also builds on the same Kimi K2.5 open-source checkpoint.
However, Cursor's secret sauce is their post-training. The continued pretraining and massive RL pipeline are what give Composer its specific developer-centric "feel".
💸 The Pricing is Absurdly Good
Despite matching or beating frontier models on benchmarks, Cursor has kept the price aggressively low.
Composer 2.5 is priced identically to Composer 2: $0.50 per million input tokens and $2.50 per million output tokens. This is a fraction of the cost of OpenAI's GPT-5.5 or Anthropic's Opus 4.7.
🚀 What's Next: The SpaceXAI Collab
Cursor is not stopping here. They announced that they are currently working with SpaceXAI to train a significantly larger model entirely from scratch. Using 10x more total compute on the Colossus 2 supercomputer, this upcoming model is expected to be a massive leap in capability.
Are you using Composer 2.5 yet? Drop your thoughts in the comments below! 👇
Top comments (0)