Cursor、コーディング向け新モデル Composer 2 を発表 Composer 2

Cursor, the company behind the AI-first code editor of the same name, has unveiled Composer 2, a new model purpose-built for agentic coding. According to Cursor, the model delivers intelligence comparable to frontier-class systems while running significantly faster, and it is available immediately inside the Cursor app.

Composer 2 is described as being trained primarily with reinforcement learning. That approach fits naturally with agentic workflows, where the model must chain tool calls, edit multiple files, run tests, and iterate based on feedback rather than simply emitting a one-shot response. The design appears to prioritize low latency as a first-class property of the developer experience, continuing the direction set by the original Composer model. In practice, that suggests Composer 2 is tuned less for long-form chat and more for the tight loops of plan, edit, run, and verify that define modern coding agents.

The announcement lands in a market that is bifurcating along strategic lines. Many AI coding tools rely on third-party frontier models from Anthropic, OpenAI, or Google, while a smaller group, including Cursor, is investing in proprietary models tailored to their product surface. Competitors such as GitHub Copilot, Windsurf, and Replit Agent are racing to deepen their own agentic capabilities, and the axis of differentiation has clearly shifted beyond raw generation quality toward responsiveness, reliability of tool use, and integration depth with the editor.

Reinforcement learning has become a near-default ingredient in recent coding-model post-training, in part because code offers verifiable rewards: tests pass or fail, programs compile or do not, diffs apply cleanly or break. Models such as DeepSeek-Coder and various open RL recipes have shown that this signal can meaningfully lift agentic performance, and Composer 2 likely benefits from a similar feedback loop, though Cursor has not detailed its training pipeline publicly. How well those gains translate into day-to-day developer productivity will depend on real-world usage rather than benchmarks alone.

Several practical questions remain open. Cursor has not, in this note, clarified pricing tiers, whether Composer 2 will be gated to paid plans, or whether it will be exposed through an external API. The framing suggests the model is meant to be experienced inside the Cursor app rather than consumed as a general-purpose endpoint, which would be consistent with the company's strategy of tightly coupling model behavior to editor context. Developers evaluating the model will want to compare it directly against Claude and GPT-class options on their own repositories, particularly on tasks involving large refactors and multi-step debugging, where speed and tool-use accuracy compound into noticeable workflow differences.