Cursor Composer 2.5 Takes Price-Performance Crown at $0.50/Task
Cursor released Composer 2.5, a proprietary coding model built on Kimi K2.5 (Moonshot) and fine-tuned on SpaceX AI's Colossus 2 cluster using custom RL with text-based feedback and 25x more synthetic tasks than Composer 2. On Cursor Bench it scored approximately 64% at roughly $0.50 per task — within 1.5 points of Opus 4.7 Max ($11/task) and GPT-5.5 extra-high ($4/task). Gemini 3.5 Flash scored ~50% at approximately 4x the cost. Composer 2.5 is priced at $0.50/M input, $2.50/M output and is Cursor-exclusive.
Why It Matters
Cost-per-task has become the dominant enterprise AI evaluation metric. Composer 2.5 sets a new price-performance floor for workhorse coding models and directly undercuts Google's flagship Flash tier at both quality and cost.