Introducing Composer 2.5

2026-05-18 · Source: Cursor Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, short

Summary

Cursor has released Composer 2.5, an enhanced version of its AI assistant, now available within the Cursor environment. This iteration significantly improves intelligence and behavior over Composer 2, demonstrating better sustained performance on long tasks, more reliable adherence to complex instructions, and improved collaborative interaction. The advancements stem from scaled training, more complex Reinforcement Learning (RL) environments, and new learning methods. Composer 2.5 is built on Moonshot's Kimi K2.5 open-source checkpoint, similar to its predecessor. Cursor is also collaborating with SpaceXAI to train a substantially larger model from scratch, utilizing 10x more compute and Colossus 2's million H100-equivalents, anticipating a major leap in capability. Composer 2.5 is priced at $0.50/M input and $2.50/M output tokens, with a faster variant at $3.00/M input and $15.00/M output tokens.

Key takeaway

For AI/ML engineering leaders evaluating advanced code assistants, Composer 2.5 offers substantial improvements in task reliability and collaborative behavior, driven by sophisticated training techniques. You should consider its new pricing tiers and the double usage offer for the first week to assess its fit for your development workflows, especially for long-running or complex coding projects. Its underlying training innovations suggest a robust foundation for future model capabilities.

Key insights

Composer 2.5 improves AI assistant performance through targeted RL, synthetic data, and advanced distributed training.

Principles

• Localized feedback improves RL credit assignment.
• Synthetic tasks scale intelligence and expose reward hacking.
• Orthogonalization optimizes distributed MoE training.

Method

Composer 2.5 uses targeted textual feedback for localized RL, dynamically generated synthetic tasks for intelligence scaling, and sharded Muon with dual mesh HSDP for efficient distributed training of MoE models.

In practice

• Apply textual feedback for specific behavior correction.
• Generate synthetic tasks to expand training data.
• Utilize HSDP with separate layouts for MoE models.

Topics

• Composer 2.5
• Reinforcement Learning
• Textual Feedback
• Synthetic Data Generation
• Distributed Training

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Cursor's Composer 2.5 matches Opus 4.7 and GPT-5.5 benchmarks at a fraction of the cost — Composer 2.5 matches top-tier coding AI performance at a fraction of the cost through extensive synthetic task training.
• Introducing Composer 2 — Composer 2 offers frontier-level coding intelligence with improved performance and competitive pricing through continued pretraining and reinforcement learning.
• Composer 2.5 and I INTERVIEWED THE CEO OF ALPHABET — Composer 2.5 delivers near-frontier coding performance at a fraction of the cost, making it a leading workhorse model for most use cases.
• Improving Composer through real-time RL — Real-time Reinforcement Learning (RL) using live user inference tokens effectively improves coding models and mitigates train-test mismatch.
• Cursor takes on OpenAI and Anthropic with Composer 2, a code-only model built to match rivals at a fraction of the cost — Specialized, cost-effective AI models can compete with generalist leaders by focusing on narrow domains.
• Introducing Claude Opus 4.8 — Claude Opus 4.8 significantly enhances agentic capabilities, honesty, and alignment, offering improved performance and new features at existing or reduced costs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Cursor Blog.