Cursor Challenges OpenAI with Composer 2, an Affordable AI Model

Composer 2: An Economical AI Model for Software Development

Cursor recently unveiled Composer 2, its artificial intelligence model dedicated to software development, with very competitive pricing. The cost is set at $0.50 per million input tokens and $1.50 per million output tokens, representing substantial savings compared to the rates of Claude Opus 4.6, which are $5.00 and $25.00 respectively, and those of GPT-5.4, at $2.50 and $15.00.

The model stands out for its specialization in code, achieving a score of 61.3 on the CursorBench, an internal benchmark by Cursor. This performance marks a significant improvement over Composer 1.5, which scored 44.2. Composer 2 thus positions itself as a serious competitor against Claude Opus 4.6, which has a score of 58.2, and GPT-5.4 Thinking, with 63.9.

Cursor's Strategy: Independence and Competitiveness

Creating its own model is a strategic decision for Cursor, which seeks to free itself from its dependence on Anthropic and OpenAI. The latter offer models at fixed rates, limiting Cursor's pricing flexibility. By developing Composer 2, Cursor aims to provide a more affordable alternative while remaining competitive in the market.

Details and Pricing of Composer 2

Composer 2 is integrated into the Cursor platform and is also available in the alpha version of the new "Glass" interface. Pricing starts at $0.50 per million input tokens and $2.50 per million output tokens. A faster version is offered, with costs of $1.50 and $7.50 per million tokens, while maintaining the same level of intelligence, according to Cursor.

Pricing Comparison per Million Tokens

• $0.50 / $2.50: Standard version of Composer 2
• $1.50 / $7.50: Accelerated version
• $5.00 / $25.00: Anthropic API rates for all context lengths
• $2.50 / $15.00: OpenAI rates for short context; $5.00 / $22.50 for long context

In terms of cost per token, Composer 2 proves to be a more economical option than Claude Opus 4.6 and GPT-5.4, even in its accelerated version.

Improvements through Reinforcement Learning

Cursor attributes the improvements of Composer 2 to ongoing enhanced pre-training, which serves as the foundation for reinforcement learning. This process focuses on complex coding tasks requiring numerous individual actions.

Results show a notable progression on the CursorBench, where Composer 2 achieves 61.3, compared to 44.2 for Composer 1.5. The model also displays increased performance on Terminal Bench 2.0, a benchmark for agent-based tasks in the terminal, as well as on SWE-bench Multilingual, which evaluates software engineering skills across multiple programming languages.

Benchmark Results

• Terminal Bench 2.0
• Optimized Terminal Bench 2.0
• SWE-bench Multilingual

Scores for Terminal Bench 2.0 vary depending on the agent and parameters used. For Claude Opus 4.6, a score of 58.0 is reported, with an optimized version reaching 65.4. GPT-5.4 Thinking has only one published value for Terminal Bench.

Challenges and Outlook for Cursor

Cursor is positioning itself directly against Anthropic and OpenAI, which continue to develop increasingly sophisticated AI models for software development. According to Bloomberg, Cursor has over one million daily users and around 50,000 enterprise clients. The company is also considering a new funding round, hoping to reach a valuation of $50 billion.

However, Cursor must navigate a complex landscape. While it develops its own model, the platform continues to support models from OpenAI and Anthropic, meaning it is competing with its own suppliers. This dependence limits its maneuverability in terms of pricing and performance.

Anthropic, with its Claude Code model, is adopting an aggressive strategy in the coding market. Cursor estimates that a Claude Code subscription at $200 per month could lead to actual computing costs of around $5,000. This highlights the structural challenge Cursor faces: by relying on third-party models, it must pay computing costs that suppliers can subsidize for their own products.

This context leaves little room for maneuver for Cursor. Consumer subscriptions are already at negative margins, supported by enterprise contracts. The long-term risk is that as AI coding agents become more efficient, users may turn away from IDEs to interact directly with the systems of model providers.