What Is Cursor Composer 2.5? (May 2026)
What Is Cursor Composer 2.5? (May 2026)
On May 18, 2026, Cursor shipped Composer 2.5 — the latest iteration of its in-house agentic coding model. Same Kimi K2 lineage as Composer 2, but with substantially heavier post-training tuned to Cursor’s exact agent harness. Here’s what it is, what changed, and why it matters.
Last verified: May 23, 2026
The 30-second summary
| Details | |
|---|---|
| Vendor | Cursor (the AI code editor company) |
| Released | May 18, 2026 |
| Base model | Kimi K2 (Moonshot AI’s 1T-class MoE) |
| Customization | Heavy Cursor-specific post-training |
| Pricing (standard tier) | $0.50 input / $2.50 output per Mtok |
| Context window | 256K tokens |
| Default in | Cursor Pro+ and Ultra (Agent + Auto modes) |
| Best for | Multi-file agentic edits inside Cursor |
| Successor to | Composer 2 (released Q1 2026) |
The history — Composer 1 → Composer 2 → Composer 2.5
Composer 1 (2024-2025): Cursor’s first attempt at an in-house coding model. Custom architecture, smaller scale, mostly used for fast single-file edits. Replaced as soon as Cursor had access to better base models.
Composer 2 (Q1 2026): Switched to Kimi K2 as the base model — Moonshot AI’s 1T-class MoE that became the default open-weight reasoning model in early 2026. Cursor added a meaningful post-training pass to make K2 work well in Cursor’s tool harness. (Coverage of Composer 2.)
Composer 2.5 (May 18, 2026): Same K2 base. Much heavier post-training. Focused on the three areas that mattered most from user feedback on Composer 2:
- Multi-file edit consistency.
- Long agent loop stability (50+ tool calls).
- Better behavior with Cursor’s specific features (Bugbot, Background Agents, Automations).
The big strategic point: Cursor is committed to having its own default model. Each iteration gets closer to “frontier quality at a fraction of frontier cost.”
What’s actually new in 2.5 (vs 2)
From Cursor’s changelog and user reports:
- Multi-file edit consistency. When Composer is editing 6+ files at once, 2.5 maintains naming consistency, type-signature compatibility, and import-order stability much better than 2.0.
- Long agent loops. Composer 2.5 can run 50+ tool calls in a single agent task without “losing the plot.” Composer 2 typically degraded around 25-30 calls.
- Bugbot integration. Composer 2.5 is the default model for Cursor’s Bugbot code-review system at all effort levels.
- Background Agents. Composer 2.5 is specifically post-trained for the long-running, low-supervision agent use case Cursor introduced in Q1 2026.
- Automations. Improvements to Cursor Automations (May 19, 2026 changelog) take advantage of Composer 2.5’s better stability.
Nothing about the user interface changes. If you use Cursor day-to-day, you’d notice “agent mode just got better at multi-file work.”
What Composer 2.5 is not
Important to be precise about limits:
- Not a frontier model. GPT-5.5 and Claude Opus 4.7 still beat Composer 2.5 on hard algorithmic problems and very long-context work.
- Not standalone. Composer 2.5 only makes sense inside Cursor. The Cursor-specific post-training is the whole point — using K2 directly outside Cursor is fine and cheaper, but you lose the harness tuning.
- Not a new architecture. Same Kimi K2 1T-class MoE base. The improvement comes from post-training, not from a new model family.
- Not free for non-Pro users. Hobby (free) tier gets limited Composer access. Pro ($20) gets included usage. Pro+ ($60) and Ultra ($200) get generous Composer 2.5 access plus parallel multi-model agent runs.
Pricing in context
Composer 2.5 keeps Cursor’s pricing structure intact:
| Tier | Price | What you get for Composer 2.5 |
|---|---|---|
| Hobby (free) | $0 | Limited Composer 2.5 calls per month |
| Pro | $20/mo | Generous Composer 2.5 in Agent + Auto |
| Pro+ | $60/mo | Composer 2.5 as default + Opus/GPT escalation |
| Ultra | $200/mo | Parallel multi-model runs (Composer 2.5 + Opus + GPT-5.5 + Gemini) |
| Teams | $40/user/mo | Pro features + admin |
| Enterprise | Custom | All features + SSO + audit |
For users on usage-based billing, Composer 2.5 is $0.50/$2.50 per million tokens (input/output) on the standard tier — roughly 10x cheaper than GPT-5.5 ($5/$30) and 30x cheaper than Opus 4.7 ($15/$75).
Real-world cost example
A typical “build a small feature across 6 files, run tests, fix errors, ship a PR” agent task consumes roughly 200K input tokens + 50K output tokens.
| Model | Cost |
|---|---|
| Composer 2.5 | ~$0.23 |
| Kimi K2 base | ~$0.25 |
| GPT-5.5 | ~$2.50 |
| Claude Opus 4.7 | ~$6.75 |
Composer 2.5’s economic advantage is exactly why Cursor can keep including it in flat-fee tiers. If Cursor had to pay frontier pricing for every agent loop, the $20 Pro tier would not be viable.
Why this matters strategically
Three things to track:
1. AI code editors are model-vendor-aware now. Cursor, Windsurf, Replit Agent, and Kiro have all built default models or aggressive Composer-style post-training. The era of “just call OpenAI’s API” is over for code editors. Each is differentiating on their own default + frontier escalation.
2. Moonshot AI’s Kimi K2 is becoming the open-weight default for coding. Composer 2.5 is the highest-profile commercial product built on K2. The fact that a $20/month consumer product uses K2 as its base validates Moonshot’s strategy and puts pressure on Meta (Llama) and Mistral.
3. The “frontier escalation” pattern is the future of agent UX. Cursor Auto mode defaults to Composer 2.5 and escalates to Opus 4.7 or GPT-5.5 only when needed. This is the dominant agent pattern in May 2026 — don’t pay for the smartest model on every call, pay for the smartest model when the task actually requires it.
Verdict
Composer 2.5 is an incremental but meaningful upgrade — the right defaults, the right pricing, the right tool-harness tuning. If you use Cursor, you’re already getting Composer 2.5 in Auto and Agent modes; no action needed.
The big-picture story is that Cursor is winning the “AI code editor” category partly because it’s stopped relying on frontier model vendors for its default experience. Composer 2.5 is the latest evidence that owning a default model is a real competitive moat, not just a cost optimization.