Cursor Ships Composer 2.5 - Frontier Coding Benchmarks at 50 Cents per Million Tokens
0:00 / 1:01
News
Cursor Ships Composer 2.5 - Frontier Coding Benchmarks at 50 Cents per Million Tokens
calendar_today Date:
schedule Duration: 1:01
visibility Views: 2
database
Summary Report
Cursor has launched Composer 2.5, an in-house coding model built on Kimi K2.5 with 25x more synthetic training tasks. It matches Opus 4.7 and GPT-5.5 on benchmarks, far cheaper.
- 01. Composer 2.5 is Cursor's new in-house coding model, replacing Composer 2 as the default agent in their AI editor.
- 02. It's built on Moonshot's open-source Kimi K2.5 checkpoint, with 85% of compute spent on Cursor's own reinforcement learning.
- 03. Trained on 25 times more synthetic coding tasks, it scores 79.8% on SWE-Bench Multilingual and 63.2% on CursorBench v3.1.
- 04. Pricing starts at 50 cents per million input tokens on the standard tier - well below frontier lab rates.
- 05. Cursor is doubling included usage for the next week and signalled a much larger model is being trained from scratch with xAI.
Cursor has launched Composer 2.5, a new coding model that powers their AI editor and delivers frontier-level performance at significantly lower costs. The model is built on Moonshot's open-source Kimi K2.5 checkpoint, but Cursor invested 85% of their compute budget into proprietary reinforcement learning and post-training processes.
The training approach paid dividends in performance metrics. Composer 2.5 was trained on 25 times more synthetic coding tasks than its predecessor, Composer 2. This extensive training regime enabled the model to achieve 79.8% on SWE-Bench Multilingual and 63.2% on CursorBench v3.1, matching the performance of Claude Opus 4.7 and GPT-5.5 on these benchmarks.
Pricing represents a significant undercut of established frontier labs. Composer 2.5 costs 50 cents per million input tokens on the standard tier, with output tokens priced at $2.50 per million. Even the faster interactive tier, whilst more expensive, remains below Anthropic and OpenAI's pricing structures.
Cursor has also revealed plans for a substantially larger model developed in partnership with xAI, using ten times the compute resources of Composer 2.5. This positions the current release as preparation for more ambitious developments ahead. To encourage adoption, Cursor is doubling included usage for existing users over the next week.
