Composer 2 by Cursor

Cursor dropped Composer 2 and it's doing something no one else in the AI coding space is doing—actually making frontier-level intelligence affordable. This thing scores 61.3 on CursorBench, 61.7 on Terminal-Bench 2.0, and 73.7 on SWE-bench Multilingual. That's a massive leap from Composer 1.5's 44.2, 47.9, and 65.9. We're talking double-digit percentage improvements across the board— Composer's first continued pretraining run built a far stronger base, then they trained it on long-horizon coding tasks through reinforcement learning. The result is a model that can solve challenging tasks requiring hundreds of actions. Not your average autocomplete trick—this is actual agent-level problem solving.

Here's the kicker: $0.50 per million input tokens and $2.50 per million output tokens. There's also a faster variant at $1.50/$7.50 that still costs less than other fast models on the market. They're making fast the default option. That combination of frontier-level capability and token-efficient pricing is exactly what developers have been waiting for—something that doesn't cost a fortune to run at scale but still actually works on the hard stuff. Available now in Cursor and the early alpha of their new interface. If you're building software and not paying attention to this, you're falling behind.