Cursor Builds Its Own Code Model, Cutting Costs 90% and Sparking a Price War

Cursor announced the release of its self‑developed programming model, Composer 2, which the company claims reaches performance comparable to GPT‑5 while costing only a fraction of the price of competing services.

📊 Data First

Internal CursorBench results show:

Composer 2 (new) – 61.3 points

Claude Opus 4.6 – 58.2 points

GPT‑5.4 Thinking – 63.9 points

Composer 1 (old) – 44.2 points

The jump from the first to the second generation moves the score from 44.2 to 61.3, surpassing Anthropic’s flagship model and approaching GPT‑5.

Pricing differences are stark:

Composer 2 input: $0.50 per million tokens

Claude Opus 4.6 input: $5.00 per million tokens

Cost gap: a full ten‑fold difference

🤔 Why Build Its Own Model?

Cursor previously relied on OpenAI and Anthropic APIs, paying competitors each time a user wrote code. Anthropic’s launch of Claude Code further threatened Cursor’s user base and forced it to subsidize a rival. The company argues that this dependency is unsustainable, making self‑development the only viable path.

🎯 What Makes Composer 2 Special?

Co‑founder Aman Sanger describes the model’s “extreme restraint”: it is trained exclusively on code and reinforced only for programming tasks. It does not generate poetry, chat, or perform unrelated calculations. This narrow focus yields remarkable accuracy on complex, multi‑step coding problems.

💡 Implications for Developers

For existing Cursor users, the shift could dramatically lower operating costs. Previously, code generation relied on expensive Claude or GPT endpoints, with costs eventually reflected in subscription fees. Composer 2’s $0.50‑per‑million‑token price represents a 90 % reduction, theoretically allowing cheaper, higher‑quality service—though actual pricing changes remain uncertain.

🌊 Bigger Picture

Cursor now serves over one million daily active users and 50,000 enterprise customers, with a valuation approaching $50 billion. The move signals that AI‑focused application companies are no longer content to remain dependent on third‑party models. The competitive landscape is shifting from “who has the strongest model” to “who can deliver the lowest cost and deepest domain expertise.”

This trend is expected to accelerate as more companies follow suit, turning the AI race into a battle over pricing and specialized performance.