Overview
Cursor Composer 2.5 is a self-developed programming model officially released by Cursor on May 18, 2026. Its base is Moonshot's open-source Kimi K2.5 (1.04T parameters / 32B active MoE), with an additional reinforcement learning post-training accounting for about 85% of Cursor's total compute. At $0.10 per task (100K input + 20K output, uncached), it is currently the most affordable cutting-edge coding model, 10x cheaper than Claude Opus 4.7 and 3x cheaper than Gemini 3.5 Flash.
Key Features
- Extreme Cost-Effectiveness: Standard version $0.50/$2.50 per Mtok, effective single task cost only $0.10, with a 2x promotion in the first week
- Dual Speed Modes: Standard (cost-effective) and Fast ($3.00/$15.00, double speed) to meet different scenarios
- Deep IDE Integration: Seamlessly works with Cursor IDE's Tab completion, Agent, and multi-file editing without context switching
- Traceable Open-Source Base: Based on Kimi K2.5 open-source weights (Modified MIT), allowing the community to audit basic capabilities
- 25x Synthetic Task Training: 25 times more synthetic programming tasks compared to Composer 2, with a +11 point improvement on CursorBench v3.1
Use Cases
- Daily feature development and bug fixes
- Multi-file refactoring in Agent mode
- Cost control for high-frequency call scenarios
- Layered strategy with Opus 4.7/GPT-5.5 (simple tasks on Composer, complex tasks on flagship models)
Pros
- Single task cost as low as $0.10, the lowest among cutting-edge coding models
- Significant coding quality improvement from RL post-training (+11 points on CursorBench)
- Smooth deep IDE integration without context switching
- Open-source base is traceable, fostering a friendly community ecosystem
Pricing
Standard: $0.50 input / $2.50 output per Mtok (approx. $0.10/task). Fast: $3.00/$15.00 per Mtok (approx. $0.60/task). Available only within Cursor IDE, no standalone API. Requires Cursor Pro ($20/mo) or Business subscription.
Summary
Composer 2.5 represents a commercial benchmark for the 'open-source base + private RL post-training' approach—offering the lowest cutting-edge coding cost in exchange for IDE lock-in. It is ideal for teams already using Cursor to upgrade seamlessly; if you need API-level flexibility or multi-IDE support, you still need to pair it with Claude Code or Codex.