Cursor 3 Composer 2 Model: Fine-Tuned Kimi K2.5 Beats Opus 4.6 at 10x Lower Cost — $0.50/M Input Tokens

VentureBeat·medium signal

Cursor's Composer 2 model, shipping with Cursor 3, is a fine-tuned variant of Moonshot's Kimi K2.5 that scores 61.7 on Terminal-Bench 2.0 and 73.7 on SWE-bench Multilingual — beating Opus 4.6 on coding benchmarks while costing $0.50/M input and $2.50/M output (roughly 10x cheaper than Opus, 5x cheaper than GPT-5.4). Composer 2 Fast delivers output at $7.50/M vs GPT-5.4 Fast's $75/M. Demonstrates that fine-tuned open models can compete with frontier models on domain-specific tasks at dramatically lower cost.