Cursor releases Composer 2.5: Innovation in training methods makes AI coding agents smarter
Cursor released Composer 2.5 on May 18, 2026, based on Kimi K2.5 checkpoint training, introducing directed text feedback reinforcement learning and sharding Muon optimizer. The new model has significantly improved long-term task execution, instruction following and collaborative experience, and is priced at $0.50/M input token. This article breaks down the technical details and practical suggestions.
Core conclusion
On May 18, 2026, Cursor officially released Composer 2.5—a major upgrade to its AI coding agent model. The new version is based on the Moonshot Kimi K2.5 open source checkpoint, using directional text feedback reinforcement learning and sharded Muon optimizer training, which has significantly improved the continuous execution of long tasks, complex instruction following, and collaborative experience. Pricing $0.50/M input token, $2.50/M output token, usage doubled in the first week.
Key Points
- Release time: 2026-05-18 -Affected objects: AI coding agent users, Cursor users, automated pipeline developers
- Core changes: directed feedback RL + sharding Muon optimizer, both intelligence and usability improvements
Background and trigger events
The Cursor team officially disclosed the technical details of Composer 2.5 on the official blog. The new model is positioned as "smarter, more suitable for long-term tasks, and more enjoyable to collaborate with" - this is not only an upgrade in model capabilities, but also an evolution in the form of AI coding agent products. Composer 2.5, like its predecessor Composer 2, is based on Moonshot's Kimi K2.5 open source checkpoint fine-tuning training, but Cursor is teaming up with SpaceXAI to train a significantly larger model from scratch with 10 times the computing power.
This release also reflects a larger trend: AI coding tools no longer compete with "who can write a line of code correctly", but with "who can run the entire project." Credit allocation, behavior calibration, and reward hacker protection in long-context rollout—these issues that were originally only discussed in RL research have now become the core competitiveness of the product.
Key Impact (by Dimension)
The following table summarizes the changes and practical significance of each dimension of Composer 2.5:
| Dimensions | Change | What it means to us | Recommended actions |
|---|---|---|---|
| Intelligence | Directed text feedback RL is added to training, and credit allocation in long Rollout is more accurate | AI coding tools can handle more complex multi-step tasks, with fewer errors from beginning to end | Use Cursor as one of the preferred tools for AI coding to test long process tasks |
| Behavior | Communication style and effort calibration are optimized by RL, and the collaboration experience is improved | Interaction is no longer rigid, AI can judge when to take the initiative and when to wait | Update Composer to the latest version to experience the optimized collaboration mode |
| Architecture | Sharded Muon + dual grid HSDP, MoE training efficiency is higher | means faster iteration speed and lower inference cost | Pay attention to the subsequent release window of larger models, and the computing cost performance will continue to improve |
| Pricing | $0.50/M input, $2.50/M output | For high-frequency coding users, Composer 2.5 is more cost-effective than similar models | Try double the usage this week to evaluate whether it is worth subscribing |
| Training | 10x computing power to train larger models from scratch (in cooperation with SpaceXAI) | The next generation model is expected to make a qualitative leap | Pay attention to subsequent announcements, the current Composer 2.5 is the best version available |
Adaptation suggestions
If you are a Cursor user, here are a few actions you can get started with:
- Upgrade to latest version of Cursor: Enabled by default in Composer 2.5. New users can experience it after creating a project.
- Test long task execution: Find a coding job that you previously needed to do manually in multiple steps, and give it to Composer 2.5 in one go. The new version's improved credit distribution in long rollouts means it can better track what it's doing.
- Try complex instructions: Composer 2.5 training adds directed text feedback - you can give more vague instructions that require reasoning and watch it correct itself through "prompts" in the context.
- Take full advantage of double the dosage in the first week: Composer 2.5 doubles the dosage in the first 7 days, suitable for high-intensity testing.
Task List (Example)
- Upgrade Cursor to the latest version
- Select a pending complex coding task as a test case
- Note the behavioral changes in Composer 2.5: Is the communication style more natural?
- Record the completion rate of long Rollout (>1000 steps), compared with Composer 2
Example: Quickly verify upgrade
How to confirm Composer 2.5 is enabled:
# Cursor (Cmd+Shift+P)
# Composer: Show Version
# 2.5
#
cursor --versionMuon
Composer 2.5 RL
(Targeted RL with Textual Feedback) RL Rollout 。 Rollout token——AI 100 , 900 。 RL ""。
Composer 2.5 ,,「...」。,。""。
** Muon + HSDP** MoE(),Muon GPU -。Cursor ——。 expert/non-expert HSDP ,CP=2 EP=8 8 GPU ( 16 )。
()
Cursor、OpenAI(GPT )、Claude(Claude Code)、DeepSeek. These tools are core components of AI coding and automation workflows.
Internal link guidance
- Want to learn more complete practical use of AI Agent tools? Watch: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes
- How to add quality gates to AI automated workflows? Watch: How to add quality gates to your AI automation workflow: A practical guide from output to trustworthy results
- Real case: He used Claude to start a business in 48 hours, with a monthly income of $9,000: Claude Code 48 hours to start a business: one person + US$29 monthly fee, monthly income in 3 months $9,000
Topic hub
AI Coding Tools Hub (2026)
From Copilot pricing changes to Claude Code + DeepSeek cost-saving setups—one place to compare tools, read explainers, and follow tutorials.
Explore AI Coding Tools Hub (2026) →Monetization angle
How can you make money from this trend?
WayToClawEarn focuses on verified earn playbooks—not just news. Start from these cases.
DeepSeek + Claude Code Micro SaaS
Run multiple small products on cheap inference
Claude Code bug bounty
Productize agent skills into security services