Summary Box [In a hurry? Just read this⚡]
- Cursor launched Composer 1.5, an upgraded version of its agentic coding model focused on better performance with complex, multi-step programming tasks.
- The model received 20 times more reinforcement learning steps during post-training than its predecessor, with more compute spent on RL than on initial pre-training.
- Composer 1.5 now uses thinking tokens to plan and reason before generating code, staying fast on simple tasks while going deeper on difficult ones.
- It includes an intelligent self-summarization mechanism to handle very long contexts by recursively condensing prior information without losing critical details.
- Available immediately in the Cursor editor for all users, with noticeable improvements reported on challenging real-world coding problems and large codebases.
Cursor has released Composer 1.5, the latest iteration of its proprietary agentic coding model, delivering substantial enhancements focused on tackling intricate and multi-step programming problems more effectively.
This update introduces several key technical improvements that make Composer 1.5 better suited for real-world development scenarios where simple code completion falls short.
One of the biggest advancements is the dramatic scaling of reinforcement learning. The model underwent 20 times more RL steps during post-training compared to its predecessor.
Notably, the computational resources dedicated to this post-training phase exceeded those used for the initial pre-training of the base model, resulting in sharper reasoning and more reliable outputs on demanding tasks.
Composer 1.5 now generates thinking tokens to actively plan and reason through codebases before producing solutions.
This adaptive thinking process allows the model to stay lightning-fast on straightforward requests while investing more depth and time on complex challenges, striking an optimal balance between speed and intelligence for daily use.
To manage extended, context-heavy workflows without losing accuracy, the model incorporates an intelligent self-summarization mechanism.
When context length approaches limits, Composer 1.5 recursively summarizes prior steps and key information, preserving logical continuity and enabling sustained performance on long-running or large-scale coding projects.
Internal benchmarks conducted by Cursor on real-world programming problems show clear gains, particularly in difficult, multi-step scenarios where earlier versions showed limitations.
The table below compares core aspects of Composer 1.5 against its predecessor:
| Aspect | Composer 1 | Composer 1.5 |
|---|---|---|
| Reinforcement Learning Scale | Baseline | 20x more RL steps |
| Reasoning Approach | Standard generation | Adaptive thinking tokens + planning |
| Long-Context Handling | Basic context management | Recursive self-summarization |
| Performance on Complex Tasks | Good for moderate difficulty | Substantial leap on challenging problems |
| Speed-Intelligence Balance | Consistent but less adaptive | Fast on simple, deep on complex |
Composer 1.5 is now publicly available within the Cursor editor for all users.
Pricing and usage limits have been adjusted in subscription plans to accommodate the model’s increased capabilities, with more generous allocation provided where capacity allows.
Developers can switch to it directly for agentic workflows, benefiting from its enhanced problem-solving without sacrificing responsiveness.
This release reinforces Cursor‘s strategy of investing heavily in post-training optimization to create a coding companion that evolves continuously.
For software engineers dealing with legacy systems, architectural refactors, or multi-file changes, Composer 1.5 promises more dependable assistance on the toughest parts of the job.
Early adopters report noticeable improvements in both efficiency and output quality when facing intricate coding demands.
