Cursor launches Composer 2
Cursor released Composer 2, a code-only model it says matches OpenAI and Anthropic code LLMs at a fraction of the cost—another sign that vertical, task-specialized models are eating into generalist provider share in developer tooling. The move intensifies the race for low-cost, high-throughput code models. (the-decoder.com) (bloomberg.com)
Composer 2 ships in two paid tiers: Standard at $0.50 per million input tokens and $2.50 per million output tokens, and a Fast default tier at $1.50/$7.50 per million; Cursor’s docs also list cache‑read discounts of $0.20 (Standard) and $0.35 (Fast) per million tokens. (cursor.com) The model is deployed natively inside Cursor with model ID composer-2 and a published 200,000‑token context window, and Cursor says the model is tuned for tool use across file edits, terminal ops and its agent stack. (cursor.com) In head‑to‑head results shared by Cursor, Composer 2 outscored Anthropic’s Opus 4.6 on Terminal‑Bench 2.0 (59.8% vs 58.0%) while remaining below OpenAI’s GPT‑5.4 (75.1%), according to independent coverage of Cursor’s published comparisons. (thenewstack.io) (thenewstack.io) Cursor attributes the jump in long‑horizon coding performance to a first continued‑pretraining run plus a reinforcement‑learning technique it calls “self‑summarization,” which the company says helps retain signal across very long agent trajectories. (venturebeat.com) (venturebeat.com) Cursor’s internal benchmark suite shows sharp iteration gains — CursorBench rising to about 61.3 (from ~44.2 for Composer 1.5), Terminal‑Bench 2.0 to ~61.7, and SWE‑bench multilingual to ~73.7 — figures Cursor published alongside the Composer 2 rollout. (webdeveloper.com) (webdeveloper.com) Bloomberg and other outlets reported the model unveiling on March 19, 2026, and notes from Cursor indicate Composer 2 is appearing first inside Cursor and the early Glass interface rather than as a broadly distributed external API. (bloomberg.com) (bloomberg.com)
