Cursor Composer 2 Built on Kimi K2.5: Self-Summarization Enables Hundreds of Sequential Actions, 61.3 CursorBench (44.2→61.3), $0.50/M Input Tokens

BuildFastWithAI·high signal

Cursor revealed Composer 2 on March 19, confirming March 20 it's built on Moonshot AI's open-source Kimi K2.5 with continued pretraining and RL. Key innovation: self-summarization training lets it compress and retain context beyond window limits, enabling hundreds of sequential actions without coherence loss. Benchmarks: CursorBench 61.3 (up from 44.2), Terminal-Bench 2.0 61.7 (beats Claude Opus 4.6's 58.0, trails GPT-5.4's 75.1), SWE-bench Multilingual 73.7. Standard at $0.50/M input tokens; fast variant at $1.50/M is the default.