- Cursor just rolled out Composer 2.5, the latest version of its AI coding model, and it’s making some claims like up to 10× better cost efficiency for long, complex software engineering work.
- This new model builds on Moonshot AI’s Kimi K2.5 checkpoint and adds more reinforcement learning, a new system for targeted textual feedback, and boosted the amount of synthetic training data 25 times more than Composer 2 had.
With this release, Cursor’s taking another big step in the fast-moving world of AI for coding. Competition is heating up among companies building these software engineering assistants. The new generation of models acts more like independent agents, they navigate huge projects, understand complex context across files, and keep long programming workflows running with less hands-on support.
Cursor says Composer 2.5 delivers more reliability, better reasoning, and higher efficiency, all while cutting token costs for developers and businesses. This shift also shows that AI coding players are leaning into training their own models, not just plugging in third-party systems.
Composer 2.5 Brings Agentic Coding and Reinforcement Learning Upgrades
Composer 2.5 is Cursor’s latest AI model for serious software engineering teams. Earlier coding assistants mostly helped with autocomplete or small chunks of code. Composer 2.5 works more independently, taking on big development tasks across codebases.
It’s built on top of Moonshot AI’s Kimi K2.5, but Cursor added extra layers of reinforcement learning, more synthetic training exercises, and behavioral tweaks to boost performance. One major addition is “targeted RL with textual feedback” which means the model can get immediate corrections during a task, not just a general review at the end.
That should help the model reason better and stick to the workflow during marathon coding sessions. In real-world use, Composer 2.5 is better at catching and recovering from mistakes along the way, holding onto context through big jobs, and using the right tools as it navigates projects.
Cursor also says this version trained on 25x more synthetic task data than Composer 2. The bigger dataset simulates what software engineers really do: debugging, analyzing repositories, and handling long, multi-step processes.
Cursor says Composer 2.5 is up to 10x more cost-efficient than similar systems. For the standard plan, it’s $0.50 per million input tokens and $2.50 per million output tokens. If you want it fast, $3.00 per million input and $15.00 per million output.
That matters because these AI coding workflows eat through tokens, especially during big development jobs. Lower costs make Composer 2.5 more appealing for engineering teams and businesses planning to scale up their AI coding operations.
Introducing Composer 2.5, our most powerful model yet.
— Cursor (@cursor_ai) May 18, 2026
It's more intelligent, better at sustained work on long-running tasks, and more reliable at following complex instructions.
For the next week, we’re doubling the included usage of the model. pic.twitter.com/N87ojcXlOC
How Composer 2.5 Has Improved from Its Earlier Models
Composer 2.5 is a clear shift in Cursor’s strategy compared with old versions. Earlier Composers mainly aimed to boost productivity right in the code editor, think faster code generation, easier refactoring, and better autocomplete.
Composer 2 and the earlier versions did those jobs well, but they struggled with extended, autonomous workflows. Keeping track of big projects or managing complex, multi-step tasks was still a weak spot.
With 2.5, Cursor is aiming for real agency. This isn’t just a reactive helper anymore. Now, it’s designed to stick with longer projects, move confidently around codebases, and interact with tools more smoothly.
They also scaled training in a big way, giving Composer 2.5 a heavily expanded synthetic dataset. Cursor believes this lets Composer 2.5 deal with more realistic coding scenarios, and it shows up in better debugging and deeper reasoning.
Besides that, Composer 2.5’s all about efficiency. Earlier models mostly competed on quality and usability, but 2.5 focuses even more on affordability and the ability to run at scale. That matches a bigger shift across the industry, where businesses care just as much about token consumption and infrastructure bills as they do about how smart the model is.
Cursor also revealed it’s training an even bigger model with “SpaceXAI” and lots more computing power. They’re clearly getting ready to take on some of the biggest AI labs in coding models.
Also read: Kimi K2.5: How Moonshot AI’s New Model Is Challenging OpenAI & Google
Conclusion
Composer 2.5 really shows how fast AI coding assistants are growing up. Compared to previous versions, it brings stronger reinforcement learning, much thicker synthetic datasets, better endurance for long tasks, and sharper cost efficiency.
You can see where priorities are changing in the industry. The race is on for coding models that can run continuously on complicated projects and that remain affordable even as companies scale up.
