Cursor Unveils Composer 2: A Frontier Coding Model That Rivals Claude Opus 4.6 at One-Tenth the Price

Anysphere, the company behind AI coding editor Cursor, has unveiled 'Composer 2,' a frontier-class coding model built in-house. This code-only model surpasses Anthropic's Claude Opus 4.6 on select benchmarks while costing just one-tenth the price. An AI coding tool that once relied entirely on API providers is now going head-to-head with its own model.

A Code-Only AI That Takes On Frontier Models

"Composer 2 can't do your taxes or write poetry. It only knows code." — Aman Sanger, Cursor co-founder

Composer 2's core strategy is abandoning generality to go all-in on coding. It combines continual pretraining with long-horizon coding reinforcement learning, optimized for agentic coding workflows that handle hundreds of sequential actions.

The results speak through benchmarks.

| Benchmark | Composer 2 | Claude Opus 4.6 | GPT-5.4 Thinking |
|-----------|-----------|-----------------|------------------|
| CursorBench | 61.3 | 58.2 | 63.9 |
| Terminal-Bench 2.0 | 61.7 | 58.0 | 75.1 |
| SWE-bench Multilingual | 73.7 | 77.8 | N/A |

On CursorBench, Cursor's proprietary benchmark, Composer 2 scored 61.3 — beating Claude Opus 4.6's 58.2 by 3.1 points. It also led on Terminal-Bench 2.0 with 61.7 versus 58.0. However, it trailed Claude Opus 4.6 on SWE-bench Multilingual (73.7 vs 77.8) and showed a significant gap against GPT-5.4 Thinking on Terminal-Bench (61.7 vs 75.1).

Composer 2 doesn't categorically dominate frontier general-purpose models, but the fact that a code-only model can compete at this level is a significant achievement in itself.

The Real Weapon Is Price Disruption

More striking than the benchmarks is the pricing.

| Model | Input (per million tokens) | Output (per million tokens) |
|-------|---------------------------|-----------------------------|
| Composer 2 | $0.50 | $2.50 |
| Claude Opus 4.6 | $5.00 | $25.00 |
| GPT-5.4 | $2.50 | $15.00 |

Composer 2's input token price is one-tenth of Claude Opus 4.6 and one-fifth of GPT-5.4. Output tokens show similarly dramatic differences. Delivering comparable performance at drastically lower cost could fundamentally reshape the economics of the AI coding tool market.

Behind this pricing strategy lies a practical calculation. Anthropic's Claude Code costs approximately $5,000 per month in compute for power users, while charging them only $200. Cursor aims to structurally avoid this negative-margin model through its own in-house solution.

1 Million DAU, $2 Billion ARR Growth Trajectory

Cursor's growth metrics are equally compelling. It has surpassed 1 million daily active users, serves 50,000 enterprise customers, and achieved $2 billion in annual recurring revenue. The company is valued at approximately $50 billion.

A user base of this scale directly translates into training data quality and volume. The code and feedback generated daily by real developers fuel a flywheel of model improvement.

Beyond the Editor — Cursor's Platform Expansion

Alongside Composer 2, Cursor announced several new features. JetBrains IDE support extends the developer ecosystem beyond VS Code. Automations uses events from Slack, Linear, and GitHub as triggers for agents to automatically execute coding tasks. A new interface called Cursor Glass was also revealed.

This marks a definitive transition from a simple code editor to a full AI coding platform. Vertical integration — an in-house model, proprietary editor, and automated pipeline — is becoming Cursor's core competitive advantage.

AI Coding Tool Market Becomes a Three-Way Race

Until now, the AI coding tool market maintained a clean separation between model providers (Anthropic, OpenAI) and tool providers (Cursor, GitHub Copilot). With Cursor building its own model, this structure is beginning to shift.

Anthropic entered the coding agent market directly with Claude Code, and OpenAI operates Codex CLI. Now with Cursor developing its own model, a vertically integrated competition spanning models, tools, and platforms has begun in earnest. For developers, the options are expanding — but choosing which ecosystem to bet on is becoming an increasingly consequential decision.