Cursor releases Composer 2 at $0.50/$2.50 per 1M tokens, undercutting Claude and GPT-4 on pricing

Cursor Launches Code-Only Model to Break Pricing Dependency

Cursor released Composer 2, its second-generation code model, priced at $0.50 per million input tokens and $2.50 per million output tokens for the standard version. A faster variant costs $1.50/$7.50. Both versions undercut rival API pricing by substantial margins.

Pricing Comparison

Model	Input	Output
Composer 2	$0.50	$2.50
Composer 2 Fast	$1.50	$7.50
Claude Opus 4.6	$5.00	$25.00
GPT-5.4 (short context)	$2.50	$15.00
GPT-5.4 (long context)	$5.00	$22.50

Performance Metrics

Composer 2 scores 61.3 on CursorBench, Cursor's internal coding benchmark—a 38% jump from Composer 1.5 (44.2) and competitive with Claude Opus 4.6 (58.2), though below GPT-5.4 Thinking (63.9).

Additional benchmarks show continued improvement across multiple evaluation frameworks:

Model	CursorBench	Terminal Bench 2.0	SWE-bench Multilingual
Composer 2	61.3	61.7	73.7
Composer 1.5	44.2	47.9	65.9
Claude Opus 4.6	58.2	58.0	77.8
GPT-5.4 Thinking	63.9	75.1	N/A

Cursor co-founder Aman Sanger told Bloomberg the model was trained exclusively on code data, enabling a smaller, cost-effective architecture. "It won't help you do your taxes. It won't be able to write poems," he said.

Training Approach

Quality gains came from stronger continued pretraining followed by reinforcement learning on long-horizon coding tasks—multi-step programming challenges requiring hundreds of individual actions. This approach drove the significant benchmark improvements over Composer 1.5 and Composer 1 (38.0 on CursorBench).

Strategic Necessity for Cursor

Building its own model addresses a structural dilemma: Cursor competes directly with Anthropic and OpenAI while depending on their APIs. As long as Cursor purchases third-party models, it faces pricing constraints its competitors don't—Anthropic and OpenAI can heavily subsidize their own products.

Cursor reportedly estimates a single Claude Code subscription at $200/month generates approximately $5,000 in compute costs for Anthropic. Consumer subscriptions at Cursor currently run at negative margins, with enterprise contracts providing profitability.

With over 1 million daily users and 50,000 enterprise customers, Cursor is discussing funding at a ~$50 billion valuation. As AI coding agents improve, the risk persists that users could bypass the IDE entirely and work directly with model providers—making Composer 2 essential to Cursor's long-term independence.

What This Means

Composer 2 represents a deliberate shift toward self-sufficiency. Cursor's pricing advantage is real but performance remains competitive rather than dominant. The code-only approach is pragmatic: narrower focus enables cheaper training and faster inference. Cursor's bet hinges on whether pricing and adequate performance can retain users against providers with deeper resources and broader models. The benchmark gap with GPT-5.4 Thinking suggests room for improvement, but SWE-bench performance (73.7) demonstrates practical engineering capability.