Composer 2.5 - Cursor's self-developed Agentic programming model

What is Composer 2.5?

Composer 2.5 is Cursor's self-developed Agentic programming model. It represents a significant improvement over Composer 2 in terms of intelligence and behavioral performance. SWE-Bench Multilingual (79.8%) and CursorBench v3.1 (63.2%) It is in the same tier as Claude Opus 4.7 and GPT-5.5 in terms of core benchmarks, but its cost per mission is only about [amount missing] of its competitors. 1/10It is known as the "king of cost-effectiveness". The model is based on continuous training using the Kimi K2.5 checkpoint from Moonshot's open source, and is currently only available through the Cursor IDE and SDK.

Main features of Composer 2.5

• Long-duration task continuous operation: Deeply optimized for long-running Agent sessions, it can maintain focus during multi-step tool calls, significantly reducing issues such as hallucinations or premature termination.
• Reliable Execution of Complex InstructionsThe reliability of following complex instructions such as cross-file reconstruction, terminal command execution, and test-driven development is significantly improved compared to Composer 2.
• Effort level dynamic calibrationThe model can automatically allocate computational load according to the difficulty of the task, completing simple tasks quickly and thinking deeply about complex tasks, avoiding "idling on small tasks and lacking thought on big tasks".
• Communication style optimizationThe responses are more concise and structured, reducing unnecessary lengthy explanations and providing a clearer reasoning process when multiple files are changed.
• Improved accuracy of tool callsSignificantly reduces invalid terminal commands or redundant searches, improving code retrieval and terminal operation efficiency.
• Flexible adaptation to dual versionsOffers Standard ($0.50/$2.50 per M tokens) and Fast ($3.00/$15.00 per M tokens) versions, with the same level of intelligence, adapted for background batch tasks and interactive real-time programming respectively.

How to use Composer 2.5

• Enable in Cursor IDEUpdate Cursor to the latest stable version as of May 2026, and open the Composer panel (Cmd+I / Ctrl+IClick the model selector to switch to Composer 2.5
• Select speed gearInteractive development uses the default method Fast version(Fast response, low latency); Background Agents or batch tasks can be switched in Settings > Models. Standard version(Low cost, same level of intelligence)
• Programmatic calls via SDK:

import { Agent } from "@cursor/sdk";
const agent = await Agent.create({
 model: "composer-2.5", // Standard 版
 // model: "composer-2.5-fast", // Fast 版
 workspace: "./",
 tools: ["edit", "shell", "search", "browser"],
});

• Limiting long-duration tasksSet limits on the number of iterations and the maximum time for unattended, long-running Agent sessions to prevent models from using "shortcuts" such as caching to cheat on rewards.
• Claim your launch bonusDuring the first week of release, users will receive double the usage credit. Credit usage can be viewed in the backend.
• Multi-model routing strategy: Delegate 90% of routine tasks to Composer 2.5; only route architecture review tasks to Claude Opus 4.7, and heavy terminal shell tasks to GPT-5.5.

The core advantages of Composer 2.5

• Ultimate cost-effectivenessStandard version input $0.50/M, output $2.50/M; Fast version input $3.00/M, output $15.00/M, approximately cheaper than Claude Opus 4.7. 10~30 times.
• Frontier-level benchmark performanceSWE-Bench Multilingual 79.8% (Opus 4.7 is 80.5%, GPT-5.5 is 77.8%), CursorBench v3.1 63.2% (basically on par with Opus 4.7's 64.8% and GPT-5.5's 64.3%).
• Deep optimization at the behavioral levelIn addition to expanding the scale of training, we have also improved the calibration of communication style and engagement level. These dimensions are difficult to be fully reflected by benchmarks, but they are crucial to the actual user experience.
• Choose between two versionsStandard is suitable for background agents and batch tasks, while Fast is suitable for interactive IDEs and real-time programming. Both have the same level of intelligence.
• Double usage bonus for first launchDouble the usage allowance during the first week of release.

Comparison of Composer 2.5 with similar competing products

Application scenarios of Composer 2.5

• Multi-file level refactoringIt has a clear cost advantage and accuracy on par with cutting-edge models, making it suitable for large-scale codebase migration.
• Interactive Pair ProgrammingThe Fast version offers rapid response and is suitable for real-time IDE collaboration.
• Background scheduled tasks/cloud agentThe Standard version offers excellent value for money and is suitable for batch code review and fixing.
• Test-Driven DevelopmentLong-duration mission reliability is superior to its predecessor, and it can stably complete multiple rounds of test-repair cycles.
• Complex terminal automationTerminal-Bench 2.0 scored 69.3%, on par with Opus 4.7 (69.4%), but still slightly inferior to GPT-5.5 (82.7%) in heavy Shell scenarios.