CursorAI/Model/Composer/2
Key Features
- Cursor's own agentic coding model (continued pretraining plus reinforcement learning on long-horizon coding tasks)
- Strong on long-horizon work via self-summarization during training
- Tuned for tool use, file edits, and terminal operations inside Cursor
- Fast variant is the default in the product (same capability profile, higher throughput pricing tier)
- Announced and documented March 2026; see Introducing Composer 2
Performance Highlights
- Reported benchmark snapshot (Cursor blog, Composer vs earlier Cursor Composer generations):
- CursorBench: 61.3 (vs 44.2 Composer 1.5, vs 38.0 Composer 1)
- Terminal-Bench 2.0: 61.7 (vs 47.9 / 40.0)
- SWE-bench Multilingual: 73.7 (vs 65.9 / 56.9)
- CursorBench methodology
Model Tiers
- Standard: $0.50/M input tokens, $2.50/M output tokens
- Fast: $1.50/M input, $7.50/M output — default for interactive use; Cursor positions it as cost-competitive vs other fast models at similar speeds
Access Details
- Cursor IDE and Cursor CLI (
cursor-agent); see CLI overview - On individual plans, usage draws from a standalone Composer usage pool with included usage; team/enterprise billing noted in models and pricing
- Early access to a new UI mentioned in the blog post (Glass)
Technical Notes
- Training story: continued pretraining for a stronger base, then RL on long-horizon coding; aimed at tasks requiring many sequential actions
- Evaluation: Terminal-Bench 2.0 scoring for Cursor used the Harbor harness per Laude Institute; details in the composer-2 blog footnotes
References