Cursor Used More Compute in Post-Training Than Pretraining to Build Composer 1.5

Cursor has launched Composer 1.5, a new version of its agentic coding model, and disclosed that the compute used during post-training exceeded the compute used to pretrain the base model.

“Our new release, Composer 1.5, strikes a strong balance between speed and intelligence for daily use,” the Cursor team wrote in a blog post.

The company also revealed that Composer 1.5 was built by scaling reinforcement learning by 20x on the same pretrained model, and that the compute used for post-training ultimately surpassed the amount used to pretrain the base model.

The company says Composer 1.5 continues to show gains in coding ability as reinforcement learning is scaled. On internal benchmarks of real-world coding problems, the model surpasses Composer 1 and continues to improve, with the largest gains on harder tasks.

“In the process of responding to queries, the model generates thinking tokens to reason about the user’s codebase and plan next steps,” Cursor stated. “We find that these thinking stages are critical to the model’s intelligence.”

Composer 1.5 is designed to adjust its reasoning depth based on task difficulty. For easier problems, it responds quickly with minimal thinking. For harder problems, it takes longer to reason producing an answer.

The model also includes self-summarisation for longer-running tasks. This allows Composer 1.5 to continue working when it reaches context limits by generating summaries and resuming exploration.

“This allows the model to continue exploring for a solution even when it runs out of available context,” the team said. “We train self-summarisation into Composer 1.5 as part of reinforcement learning by asking it to produce a useful summary when context runs out in training.”

Cursor recommends Composer 1.5 for interactive use and reports that its training shows reinforcement learning for coding can be scaled with predictable improvements.

Cursor’s Composer is an AI coding model built specifically for software engineering tasks. Launched as part of Cursor’s core product suite, Composer is trained through reinforcement learning and designed to operate like an autonomous coding agent.

Cursor has also updated pricing alongside the launch of Composer 1.5. According to the company’s pricing page, Composer 1.5 is listed at $3.50 per million input tokens, $3.50 per million cache writes, $0.35 per million cache reads, and $17.50 per million output tokens. 

By comparison, Composer 1 is priced at $1.25 per million input tokens, $1.25 per million cache writes, $0.125 per million cache reads, and $10 per million output tokens.