In benchmarks such as SWE-Bench Multilingual, where it scored 79.8%, and CursorBench v3.1, where it reached 63.2%, Composer 2.5 performs on a similar level to Opus 4.7 and GPT-5.5. However, Cursor says it is significantly cheaper per task: $0.50 per million input tokens and $2.50 per million output tokens, a fraction of the costs associated with Anthropic and OpenAI models. A faster version with the same performance is available for $3.00 and $15.00, respectively.

Composer 2.5 is now available in Cursor, with further details provided in the model documentation.

Composer 2.5 achieves similar performance to Opus 4.7 and GPT-5.5 in CursorBench 3.1, but costs less than one dollar per task instead of up to eleven dollars.
Composer 2.5 achieves similar performance to Opus 4.7 and GPT-5.5 in CursorBench 3.1, but costs less than one dollar per task instead of up to eleven dollars.

Together with SpaceX and xAI, Cursor is already training a much larger successor model “from scratch” using ten times more compute on the Colossus-2 cluster, which has one million H100 equivalents.

Сonclusion:
Composer 2.5 strengthens Cursor’s position in the AI coding market by combining competitive benchmark results with much lower operating costs. Its next large-scale model, trained with SpaceX and xAI, could further intensify competition among AI coding tools.