Cursor launches Composer 2.5, built on Kimi K2.5, claiming SOTA coding performance at 1/10th the cost of frontier models.
Key Takeaways
Composer 2.5 is based on Kimi K2.5, the same base as Composer 2, with claimed benchmark parity against top-tier Anthropic and OpenAI models.
Cursor claims roughly 10x cost efficiency versus comparable frontier models.
Cursor has begun training a new model from scratch on the Colossus 2 cluster (xAI/SpaceX infrastructure), signaling a longer-term model strategy.
Full release details are available at cursor.com/blog/composer-2-5.
Hacker News Comment Review
Skepticism runs high: Composer 2 carried similar eval-backed SOTA claims that did not hold up in day-to-day usage, and at least one commenter canceled their subscription and moved to Claude Code as a result.
Commenters flagged a benchmark framing choice: Cursor compares against Opus rather than Sonnet, which some see as cherry-picking the ceiling while avoiding a more relevant mid-tier comparison.
The Colossus 2 training detail and the Kimi K2.5 base drew genuine interest as evidence Cursor is no longer purely a wrapper, though eval credibility remains the central debate.
Notable Comments
@PUSH_AX: “Composer 2 which was evaled in close competition with frontier models, spoiler alert, it wasn’t even close in practice” – direct challenge to Cursor’s benchmarking methodology.
@asar: Confirms Kimi K2.5 base and Colossus 2 training detail; notes the perception shift from “VSCode fork with no moat” to serious model developer.
