Cursor details Composer 2's training stack in a new technical report

On March 24, 2026, Cursor’s X account said it was releasing a technical report explaining how Composer 2 was trained. The announcement points developers to a deeper write-up that moves beyond the launch blog and explains the training, evaluation, and systems work behind the model.

What Cursor disclosed

According to Cursor’s March 27 technical report and March 19 launch post, Composer 2 is trained in two phases: continued pretraining on a code-heavy mix to strengthen base coding knowledge, followed by large-scale reinforcement learning to improve end-to-end agent performance. Cursor says the RL setup uses the same tools and harness as the deployed product, with tasks drawn from the kinds of ambiguous, multi-file problems developers actually ask the product to solve.

Cursor also introduced CursorBench as an internal benchmark derived from real engineering sessions. On the metrics cited in the official posts, Composer 2 scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual. Cursor positions the model as a frontier-level coding system with lower cost than prior Composer releases, listing $0.50 per million input tokens and $2.50 per million output tokens for the standard variant.

Why it matters

The announcement is notable because coding-model vendors rarely disclose much about the path from base model to agent behavior. Cursor’s report makes the case that agentic coding quality depends not just on a stronger base model, but on training in realistic tool-using environments, long-horizon tasks, and evaluation loops that look closer to production than to narrow benchmark prompts. For teams using coding assistants, the key takeaway is that model training recipes are increasingly being optimized around full workflow completion rather than isolated code generation.

Source materials include Cursor’s X post, its technical report, and the Composer 2 launch article.