Cursor has released Composer 2, the second generation of its in-house AI model for software development. The company says the model can compete with leading coding models from Anthropic and OpenAI at significantly lower cost.
The model is now available in Cursor and also in the early alpha of the new interface called “Glass.” Pricing is set at 0.50 dollars per million input tokens and 2.50 dollars per million output tokens. A faster version, which Cursor says offers the same level of intelligence, costs 1.50 dollars and 7.50 dollars per million tokens respectively and is enabled as the default option.
| Model | Price per 1M tokens, input / output | Note |
|---|---|---|
| Composer 2 | 0.50 / 2.50 dollars | Standard version |
| Composer 2 Fast | 1.50 / 7.50 dollars | Faster version with, according to Cursor, the same intelligence |
| Claude Opus 4.6 | 5.00 / 25.00 dollars | API pricing according to Anthropic, valid for any context length |
| GPT-5.4 | 2.50 / 15.00 dollars, short context; 5.00 / 22.50 dollars, long context | OpenAI pricing depending on context length |
On pure API pricing, Cursor positions Composer 2 well below Claude Opus 4.6 and also below GPT-5.4. Even the faster Fast version still comes in far below the standard token pricing of both larger competitors.
Speaking to Bloomberg, co-founder Aman Sanger said the model was trained exclusively on code data. According to him, that focus made it possible to build a smaller and cheaper model. “It won’t help you with your tax return,” Sanger said. “It won’t be able to write poems.”
Reinforcement learning on long coding tasks as the training foundation
According to Cursor, the quality improvements over the previous version are largely due to the company’s first run of continued pretraining, which provided a stronger base for the reinforcement learning stage that followed. Training is based on so-called long-horizon coding tasks, meaning programming problems that require hundreds of individual actions.
The figures published by Cursor show a particularly clear jump compared with earlier Composer versions. On CursorBench, the company’s internal benchmark for coding tasks, Composer 2 rises from 44.2 for Composer 1.5 to 61.3.
The model also shows clear gains on Terminal Bench 2.0, a benchmark for agentic tasks in the terminal, and on SWE-bench Multilingual, a benchmark for software engineering tasks in multilingual environments.
| Model | CursorBench | Terminal Bench 2.0 | Terminal Bench 2.0 optimized | SWE-bench Multilingual |
|---|---|---|---|---|
| Composer 2 | 61.3 | 61.7 | 73.7 | |
| Composer 1.5 | 44.2 | 47.9 | 65.9 | |
| Composer 1 | 38.0 | 40.0 | 56.9 | |
| Claude Opus 4.6 | 58.2 | 58.0 | 65.4 | 77.8 |
| GPT-5.4 Thinking | 63.9 | 75.1 | n/a |
Terminal Bench 2.0 is only partially comparable directly, since results also depend on the agent, harness, and settings. For Claude Opus 4.6, 58.0 is the public Claude Code score, while 65.4 is an additional optimized result published by Anthropic. For GPT-5.4 Thinking, only a single published Terminal Bench score is currently available.
Cursor needs to become less dependent on its suppliers
Cursor now competes directly with Anthropic and OpenAI, both of which are offering increasingly capable AI models for software development. According to Bloomberg, Cursor now has more than one million daily users and around 50,000 enterprise customers. The company is also reportedly in talks over a new funding round at a valuation of around 50 billion dollars.
At the same time, Cursor faces a structural dilemma. The platform still supports models from OpenAI and Anthropic and therefore competes directly with the very providers on whose technology it has so far depended. As long as Cursor relies on external models, its pricing, capabilities, and ultimately its own margins depend on companies that are serving the same customer base directly.
Anthropic in particular is gaining a dominant position in the coding market with Claude Code. Cursor reportedly estimates internally that a 200-dollar-per-month Claude Code subscription could now generate computing costs of around 5,000 dollars. That illustrates the problem clearly: if you embed a third-party model into your own software, you are buying compute at prices the model provider can afford to heavily subsidize for its own product.
That leaves Cursor with limited room to maneuver. According to the report, consumer subscriptions are running at negative margins, while enterprise contracts are the main source of profitability. On top of that, increasingly agentic models could reduce the value of the interface layer itself if users begin going directly to the model provider instead of using an intermediary platform.
ES 
EN