Cursor Composer 2.5 Becomes the Most Chosen Model in Cursor — 10x Usage Bonus

Cursor announced its proprietary agentic coding model Composer 2.5, revealing that it became the most chosen model on Cursor immediately after launch. To celebrate, Cursor gave all users 10x daily usage.

Composer 2.5 is the 3rd-generation model, officially launched on May 18, and is based on Moonshot AI's open-source Kimi K2.5. This time, Cursor explicitly named the base model in the first paragraph, explaining that it put 85% of its total compute into its own RL and post-training. It also stated that it used 25x more synthetic coding tasks compared to Composer 2.

On performance, several benchmarks were presented.

• Terminal-Bench 2.0: 69.3%, nearly tied with Opus 4.7's 69.4%, falling short of GPT-5.5's 82.7%
• SWE-Bench Multilingual: 79.8%, close to Opus 4.7's 80.5% and above GPT-5.5's 77.8%
• CursorBench v3.1 default settings: 63.2%, ahead of Opus 4.7 xhigh default (61.6%) and GPT-5.5 medium default (59.2%)

Pricing was also presented as a strength. The Standard tier is $0.50/M for input and $2.50/M for output, while the Fast tier is $3.00/M for input and $15.00/M for output. Cursor emphasized that it delivers Opus 4.7-level performance on SWE-Bench Multilingual at about 1/10 of Anthropic's pricing.

Additionally, Cursor revealed that it is training a model 10x larger from scratch together with xAI infrastructure, and stated that currently 35% of PRs merged internally are generated by autonomous agents. Composer 2.5 is only available on Cursor IDE, CLI, and web.