Cursor Composer 2.5: Frontier-Level Coding at $0.50 Per Task

Cursor’s new Composer 2.5 model scores 79.8% on SWE-Bench Multilingual — just 0.7 points behind Anthropic’s Opus 4.7 (80.5%) — while costing roughly $0.50 per task versus $7 for Opus 4.7 at its default settings.

The Open-Weight Fine-Tuning Thesis

Cursor built Composer 2.5 on Moonshot AI’s open-source Kimi K2.5 checkpoint, then fine-tuned it with 25x more synthetic training tasks than its predecessor, Composer 2. The training method is worth noting: Cursor uses a “feature deletion” technique where an agent is given a real codebase, deletes code until specific tests break, and must reimplement the missing functionality — with test success as the reward signal. This grounds synthetic data in real engineering problems rather than toy benchmarks.

The result is a model that doesn’t need frontier-scale compute to produce frontier-grade coding outcomes. On CursorBench v3.1 — Cursor’s harder internal benchmark — Composer 2.5 hits 63.2% accuracy while Opus 4.7 at its “xhigh” setting reaches ~62% at over 10x the cost per task. On Terminal-Bench 2.0, Composer 2.5 scores 69.3%, essentially tying Opus 4.7’s 69.4%.

What the Cost Gap Actually Means

The per-task economics matter more than benchmark numbers for teams running agentic coding workflows at scale. Cursor’s standard-tier pricing sits at $0.50 per million input tokens and $2.50 per million output tokens — the fast tier (default for interactive use) runs $3/M input and $15/M output. At benchmark task complexity, that translates to roughly $0.50 per automated coding task, compared to $7 or more for Opus 4.7 and approximately $2.20 for GPT-5.5.

For a platform team running hundreds of agentic code reviews, migration tasks, or test-generation jobs per week, that differential compounds quickly. The efficiency argument isn’t purely about raw capability — it’s about whether frontier-equivalent output can be sustained economically at the volume that agentic workflows demand.

The Pattern

Fine-tuned open-weight models are closing the gap with closed-source frontier labs on coding benchmarks. Stack Archive has tracked this emerging pattern through Cursor’s earlier parallel-fleet architecture — Composer 2.5 is the clearest data point yet that the gap is now narrow enough to matter for production decisions.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.