Cursor released Composer 2 — a frontier coding model at $0.50 per million tokens

Cursor has presented Composer 2 — a new coding model that demonstrates frontier-level performance and transforms the economics of AI-powered coding. The model is available now in the Cursor editor and is accompanied by a technical report on its training.

Frontier-level on benchmarks

Composer 2 has achieved significant improvements across all major coding benchmarks. On its own CursorBench 2, the model scored 61.3 points — 38% better than Composer 1.5. Even more impressive results came on SWE-bench Multilingual, where Composer 2 achieved 73.7 versus 65.9 for the previous version. On Terminal-Bench 2.0, the result was 61.7.

• CursorBench: 61.3 (+38% vs Composer 1.5)
• Terminal-Bench 2.0: 61.7
• SWE-bench Multilingual: 73.7 (+13% vs Composer 1.5)

These improvements are based on the first cycle of continued pretraining — when the model is retrained on enhanced data. This provides a stronger foundation for subsequent reinforcement learning training, which teaches the model to solve complex multi-step tasks.

A price that changes the rules

Composer 2 costs $0.50 per million input tokens and $2.50 per million output tokens. For comparison: this is one of the most affordable options on the market at this level of quality. Cursor has also released a faster version of the model with the same performance — at $1.50 for input and $7.50 for output tokens. The faster variant is now the default, and its price is lower than other fast-model competitors.

Such a combination — high competence plus low cost — makes Composer 2 an optimal choice for many development workflows. For individual plans, Composer usage is included in a separate pool with a generous limit.

How this was achieved

The improvements are based on a new training approach. First, Cursor conducted continued pretraining on the best coding data, creating a stronger starting point for further optimization. Then the model was trained on long-horizon tasks through reinforcement learning — tasks requiring hundreds of actions and decisions. This is what allowed Composer 2 to handle truly complex coding scenarios, including multi-file projects and architectural decisions.

What this means

The AI-coding market is entering a new period — where quality and cost are no longer traded off against each other. For developers, this means powerful coding tools are becoming more practical and accessible. For companies like Cursor, this signals that the boundary between standard and frontier now runs along the speed of learning and economic efficiency, rather than absolute model performance.