Cursor Unveils Composer 2: A Code‑Focused Model Priced at a Fraction of GPT‑5

Cursor announced the second‑generation self‑built programming model Composer 2. In internal benchmarks the model’s score rose from 44.2 in the first generation to 61.3, surpassing Anthropic’s Claude Opus 4.6 (58.2) and approaching OpenAI’s GPT‑5.4 Thinking (63.9).

Composer 2’s training strategy is deliberately narrow: it uses reinforcement learning exclusively on code data. Consequently, the model does not generate poetry or perform tax calculations, but it delivers remarkably high accuracy on complex programming tasks that require hundreds of steps.

Pricing is a key differentiator. The standard Composer 2 costs $0.50 per million input tokens, whereas Claude Opus 4.6 charges $5.00 for the same volume. Even the high‑throughput “Fast” variant remains far cheaper than competing offerings.

For Cursor, the move to a proprietary model addresses both performance and survival concerns. Previously the company depended heavily on OpenAI and Anthropic APIs, paying steep fees while competitors subsidized prices to attract users. With over one million daily active users, 50,000 enterprise customers, and a valuation near $50 billion, Cursor’s launch of Composer 2 represents a full‑stack shift from editor to underlying model, aiming to seize control of the second half of the AI‑coding market.

Recent market trends show Claude‑code gaining momentum while other products lose visibility, highlighting Cursor’s niche position—neither a dominant application channel nor a pure model provider. By focusing on speed and low cost, Cursor attempts a breakthrough in the middle‑layer product space. Some observers argue that token pricing should reflect effectiveness; the analogy likens this to a power plant not caring how a light bulb saves electricity, but the bulb seizing the opportunity.

As a deep Cursor user, I hope they can explore a middle‑layer product path.