In a major strategic move to reduce its dependency on external AI providers, Cursor (Anysphere Inc.) officially launched Composer 2 on March 19, 2026. The new in-house model is specifically designed for “agentic” coding—tasks where the AI autonomously navigates, edits, and tests code across large repositories—while significantly undercutting the pricing of rivals like OpenAI and Anthropic.
The Specialization Advantage
Unlike general-purpose models (GPT-5.4 or Claude 4.6), Composer 2 was trained exclusively on coding-related data. This narrow focus allowed Cursor to build a smaller, more efficient model that rivals the “intelligence” of much larger systems in software engineering tasks.
- “It won’t write poems”: Cursor co-founder Aman Sanger highlighted that the model’s lack of general knowledge (like doing taxes or writing poetry) is what makes it so lean and cost-effective for developers.
- Agentic Prowess: The model is optimized for “long-horizon” tasks, meaning it can solve complex problems that require hundreds of individual actions, such as codebase-wide refactoring.
Pricing: A Radical Shift in Economics
The launch of Composer 2 effectively resets the pricing for high-end AI coding. Compared to its predecessor (Composer 1.5), the new model is 86% cheaper.
| Model Variant | Input (per 1M tokens) | Output (per 1M tokens) | Key Feature |
| Composer 2 Standard | $0.50 | $2.50 | High-precision agentic coding |
| Composer 2 Fast | $1.50 | $7.50 | Same intelligence, 4x speed |
| Claude Opus 4.6 | $5.00 | $25.00 | Market benchmark (for reference) |
| GPT-5.4 Thinking | $2.50 – $5.00 | $15.00 – $22.50 | Leading reasoning model |
Performance Benchmarks
Cursor published results showing that Composer 2 has moved into the “frontier” tier, outperforming the recently released Claude Opus 4.6 on several key developer metrics.
- CursorBench: Climbed from 44.2 (v1.5) to 61.3, placing it just behind GPT-5.4.
- Terminal-Bench 2.0: Scored 61.7, beating Claude Opus 4.6 (58.0) in its ability to execute commands in a CLI.
- SWE-bench Multilingual: Reached 73.7, demonstrating strong capabilities across diverse programming languages.
Key Technical Features
- 200,000-Token Context: Allows the model to ingest and understand large chunks of a codebase or long documentation in a single prompt.
- Self-Summarization: A unique training technique that helps the model compress data for extremely long-running tasks, preventing it from “forgetting” the original goal.
- Cache-Read Discounts: Developers can save even more with $0.20–$0.35 per million tokens for repeated prompts, rewarding iterative development.
The “Vibe Shift”: Toward Self-Driving Codebases
The release of Composer 2 marks a fundamental shift in Cursor’s product philosophy. While the tool began as a “collaborative editor,” leadership is now prioritizing “Priority Zero”: building the world’s best autonomous coding model. This move targets “Grind Mode,” a future feature where hundreds of agents could work simultaneously on a project without human intervention.
“The future of software development does not involve writing lines of code,” Cursor leadership reportedly told employees. “It involves directing the agents that do.”
