Cursor Composer 2.5

Overview

Cursor Composer 2.5, released May 18, 2026, is the next iteration of Cursor's proprietary coding model. It is built on the same Moonshot Kimi K2.5 open-source checkpoint that Composer 2 uses, with Composer 2.5-specific training improvements rather than a new base model. Cursor describes the upgrade as a "substantial improvement in intelligence and behavior" with the headline change being calmer collaboration on long-running tasks.

Pricing matches Composer 2 on the Standard variant at $0.50 per 1M input tokens and $2.50 per 1M output tokens. The Fast variant — which has been the default model surface inside Cursor since Composer 2 — is repriced to $3.00/M input and $15.00/M output. All users get double usage credits in the first week of launch.

What's New

Same Kimi K2.5 base, new training stack

Composer 2.5 keeps the open-source base that Cursor first adopted for Composer 2 (Moonshot's Kimi K2.5 checkpoint) and rebuilds the training stack on top of it. The launch post highlights training improvements such as sharded Muon optimization and dual-mesh HSDP, with an optimizer step time of about 0.2 seconds on the 1T-parameter model. This is an iteration on Composer 2's foundation, not a new architectural shift.

25× more synthetic training tasks

The largest training-side change is volume: Composer 2.5 was trained on roughly 25× more synthetic coding tasks than Composer 2. Cursor uses these synthetic tasks to expand coverage of complex RL environments and to support new learning methods that Composer 2's smaller synthetic dataset could not. Cursor attributes much of the "better at sustained work on long-running tasks" claim to this expansion.

Targeted textual feedback in RL

Beyond more synthetic tasks, Composer 2.5 introduces targeted textual feedback during reinforcement learning — the model receives written critique on its actions rather than only numeric reward signals. Cursor presents this as a key driver behind the improved instruction-following and effort calibration that 2.5 advertises.

Calmer, better-calibrated agent behavior

Cursor explicitly frames 2.5 as "more pleasant to collaborate with." Post-training prioritized stronger instruction-following on complex, multi-step prompts and improved effort calibration, so the model spends more reasoning on harder tasks and less on simple ones rather than running at full effort by default. The launch post does not advertise a speed regression on simple tasks.

Separate SpaceXAI larger-model partnership (forward-looking)

Cursor's launch post also notes a separate effort: together with SpaceXAI, Cursor is training a significantly larger model from scratch using 10× more total compute. The launch post presents this as forward-looking context about Cursor's future model training, not as a confirmed Composer 2.5 training input. Treat the 10× number as a roadmap signal, not a Composer 2.5 specification.

Pricing and default surface

Standard pricing is held flat at $0.50/M input and $2.50/M output. The Fast variant is repriced to $3.00/M input and $15.00/M output. Fast remains the default model surface inside Cursor, the same as in Composer 2. The double-credit week-one promotion applies to both variants.

Migration Guide

From Composer 2

The launch post does not mention an API rename or breaking change to surface; Cursor swaps the underlying model when you opt into 2.5. Practical advice:

• Re-run any internal evals that previously baselined Composer 2 — Cursor calls out substantial behavioral changes (instruction-following, effort calibration, textual-feedback-trained), not just raw capability gains, so behavior on your task set may shift in ways that aren't benchmark-visible.
• Fast remains the default, but its token price is higher than Composer 2 Fast; review model selection and usage settings before migrating cost-sensitive workflows.
• Use the double-credit week-one window for migration testing rather than incremental adoption.

From Composer 1.5

The 1.5 → 2 → 2.5 path is large. Composer 2 was already a major step up on CursorBench (44.2 → 61.3) and Terminal-Bench 2.0 (47.9 → 61.7). Composer 2.5 is described qualitatively as a further "substantial" jump on top of that, and Cursor's launch post embeds a benchmark comparison table image — see the media attached to this page for the official chart.

API and channel access

Composer 2.5 is generally available in Cursor today. Cursor's launch post does not announce a separate third-party API for Composer 2.5 at launch — programmatic use still happens through the Cursor editor or its existing integrations.

Pricing & Plans

Composer 2.5 is billed per token via the Cursor IDE. Cursor's underlying subscription tiers (Hobby / Pro / Business) determine monthly credit allocations.

Cursor Pro starts at $20/month and includes monthly Composer credits; heavy usage is billed per the table above. Composer 2.5 changes Fast variant pricing versus Composer 2, so verify current Cursor model pricing on the official pricing page before setting production budgets.

Best For

• Cursor teams that hit Composer 2's ceiling on multi-hour, multi-file refactoring or long agentic loops.
• Engineering orgs comparing IDE-native coding agents against external API-based agents like Claude Code or Codex CLI.
• Developers who prefer the Fast variant's response speed and can absorb the higher per-token cost.
• Teams that already standardized on Cursor Pro / Business and want the latest Composer model without changing tooling.
• Solo developers running long-form agentic experiments who can take advantage of the week-one double-credit window.

FAQ