Descriptions:
Matthew Berman reviews Cursor’s newly released Composer 2.5, the latest in-house coding model from Cursor built on the Kimi open-source model family. Despite being positioned as a minor incremental release, Berman argues the model is a meaningful inflection point: on Cursor’s internal CursorBench evaluation — which plots models on axes of cost per task versus benchmark score — Composer 2.5 achieves performance close to frontier models at an estimated $0.50 per task, compared to roughly $11 per task for Anthropic’s Opus 4.7 and approximately $4 for OpenAI’s GPT 5.5 Extra High. The model is currently exclusive to the Cursor IDE and is not available through any external API.
Berman contextualizes the release against Google’s simultaneous launch of Gemini 3.5 Flash, arguing that a broader convergence is underway around what he calls “workhorse-class” models — fast, cheap, and capable enough for the majority of production coding tasks. The video includes a segment from his conversation with Google CEO Sundar Pichai at Google IO, in which Pichai explains that serving billions of users across Search and Gemini makes cost efficiency a strategic necessity, not just an optimization.
The central thesis is that raw benchmark rankings matter less than price-performance ratio for most real-world deployments. Berman argues that the vast majority of coding use cases don’t require absolute frontier performance, and that Composer 2.5’s cost profile makes it the practical default choice for developers and teams operating under realistic budget constraints.
📺 Source: Matthew Berman · Published May 26, 2026
🏷️ Format: Review
