On March 19, 2026, AI coding platform Cursor (developed by Anysphere, valued at $29.3 billion) officially announced its proprietary coding-specialized model, "Composer 2."
Composer 2 is the third-generation model following the original Composer in October 2025 and Composer 1.5 in February 2026. It delivers frontier-level coding performance at a significantly lower cost.
Benchmark results
Composer 2 achieved substantial score improvements across major coding benchmarks.
| Benchmark | Composer 1 | Composer 1.5 | Composer 2 | Claude Opus 4.6 | GPT-5.4 |
|---|---|---|---|---|---|
| CursorBench | 38.0 | 44.2 | 61.3 | 58.2 | 63.9 (Thinking) |
| Terminal-Bench 2.0 | — | 47.9 | 61.7 | 58.0 | 75.1 |
| SWE-bench Multilingual | — | 65.9 | 73.7 | — | — |
On Terminal-Bench 2.0, it recorded 61.7, surpassing Claude Opus 4.6 (58.0). Meanwhile, a gap remains with GPT-5.4 (75.1), and model competition across the industry continues.
Note that benchmark scores were measured using the Harbor evaluation framework based on a snapshot of Cursor traffic as of March 18, 2026.
Technical breakthrough: Self-Summarization
Composer 2's biggest technical innovation is a new training method called Compaction-in-the-Loop RL (self-summarizing reinforcement learning).
Conventional challenges
The processing history (trajectory) of AI agents expands faster than model context windows can accommodate. Conventional approaches addressed this issue using the following methods:
- Prompt-based summarization — Compressing intermediate context using another model or prompt
- Sliding window — Simply discarding older context
Both shared the common weakness of losing critical information.
Composer 2's approach
Composer 2 incorporates the summarization process itself into the RL training loop.
- During training: When generation reaches a token length threshold, the model itself compresses the context to approximately 1,000 tokens (a dramatic reduction from over 5,000 tokens in conventional methods)
- Reward design: The final reward is applied to the entire chain (agent response + self-summarization), reinforcing good summaries and penalizing summaries that lose information
- Result: Processing continues while preserving vital information even across long-horizon tasks spanning hundreds of steps
According to Cursor's research, self-summarization reduced compression errors by 50%.
Architecture and training
Composer 2 is a coding-specialized model adopting a Mixture-of-Experts (MoE) architecture.
- Continual pre-training: For Composer 2, proprietary continual pre-training was conducted for the first time to strengthen the RL base model
- Reinforcement learning: Trained within real development environments while utilizing developer tools such as semantic search, file editing, and terminal commands
- Long-horizon task capability: Capable of solving complex coding tasks requiring hundreds of sequential actions
Pricing
Composer 2 achieves an approximate 86% cost reduction compared to the previous generation.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Composer 1.5 | $3.50 | $17.50 |
| Composer 2 | $0.50 | $2.50 |
| Composer 2 (Fast) | $1.50 | $7.50 |
Offering frontier-level performance at this price tier is a major differentiator against general-purpose models from OpenAI and Anthropic.
Alpha version of the new interface
Concurrently with the release of Composer 2, Cursor also released an early alpha version of a new interface. This further evolves the agent-first UI introduced in Cursor 2.0 in October 2025, making parallel execution of multiple agents and coordination with background agents more intuitive.
Conclusion
Composer 2 is an important milestone demonstrating that coding-specialized models can rival or even surpass general-purpose LLMs.
The new training paradigm of self-summarizing RL is a promising approach to the common challenge of long-context processing in AI agents, and it has the potential to influence overall AI development going forward, beyond just the coding domain.
Composer 2 is available inside the Cursor editor starting today.
References:
