Cursor Releases "Composer 2.5" — Opus-Class Coding at About 1/10 the Cost

What happened: The AI code editor Cursor released its own coding agent, "Composer 2.5." Built on China's Moonshot open-source "Kimi K2.5," it reportedly scored 79.8% on the multilingual coding benchmark SWE-Bench Multilingual, reaching accuracy on par with Claude Opus 4.7 and GPT-5.5.

Why it matters: API pricing is $0.50 input / $2.50 output (per million tokens). Against the $15 of Opus 4.7 and GPT-5.5, that's about 1/10, making it possible to run frontier-class autonomous coding dirt cheap. (Pricing as of 2026-05-31.)

What it means: Even solo developers can have an agent edit multiple files at once without worrying about cost. The assumption that "frontier models are expensive" is breaking down, lowering the bar for mass-producing your own tools and apps.