Cursor’s New Model: Composer 2.5, Capabilities, and API Access

Direct answer

Cursor’s model documentation identifies Composer 2.5 as its own agentic coding model. Cursor’s changelog says it became available on May 18, 2026. That identifies the model and date; it does not establish that Composer 2.5 is the newest model available across Cursor’s entire model picker. (cursor.com)

For developers asking about an API, the key distinction is that Cursor documents programmatic access to Cursor agent workflows, not a standalone Composer prompt-in, completion-out API. Its Python SDK explicitly shows selecting Composer 2.5 for a local agent run. (cursor.com)

What Composer 2.5 is

Cursor describes Composer 2.5 as an agentic model tuned for tool use, file edits, and terminal operations. Its documentation lists a 200k context window and says the model builds on Composer 2 with improvements to long agentic tasks, effort calibration, tool selection, intent understanding, and reliability. These are Cursor’s product descriptions, not a guarantee of success on every task or codebase. (cursor.com)

“Composer” can refer to both a model and Cursor’s agent workflow. The workflow includes the surrounding tools and actions; the model is one part of that system. This matters for API decisions: Cursor says its SDKs and Cloud Agents API run agent workflows with workspace context, tools, commands, and edits, rather than exposing a standalone inference or chat-completions API. (cursor.com)

What is known about its origin—and what is not

Cursor’s Composer 2.5 announcement says it uses the same open-source checkpoint as Composer 2: Moonshot’s Kimi K2.5. That shared starting point does not mean the Cursor model is identical to Kimi. (cursor.com)

In a separate report about Composer 2, TechCrunch quoted Cursor’s Lee Robinson estimating that about one quarter of the final model’s compute came from the base model and the rest from Cursor’s training. That estimate concerns Composer 2; it should not be treated as a measured compute breakdown for Composer 2.5. The report provides no benchmark names, scores, or methodology for Robinson’s statement that Composer 2 performed differently from Kimi. (cursor.com)

How to decide whether it fits your coding work

Composer 2.5 is a candidate to evaluate for multi-step work involving tool use or coordinated file edits. Cursor describes it as strong on long-horizon tasks, but the product description alone does not show that it is the best choice for your repository.

For a practical comparison:

1. Check access and workflow fit. Confirm Composer 2.5 is available in the Cursor setup you plan to use.
2. Choose representative tasks. Try work you actually do, such as a contained refactor or adding tests, with clear acceptance criteria.
3. Keep conditions consistent. Give each candidate the same repository context, task, tools, and criteria.
4. Inspect outcomes. Review the diff, run relevant tests, and check for unwanted changes. Track time or spend if those affect your decision.

The sources cited here do not establish a universal winner across coding tasks. Use your own representative tasks to decide.

Can you use Composer 2.5 through a production API?

Not as a standalone model-inference API in the Cursor documentation cited here. Cursor describes its APIs and SDKs as ways to run agent workflows, not as a raw model endpoint. (cursor.com)

The Python SDK does show Composer 2.5 selected for a local agent run. Its example uses the model ID composer-2.5. The SDK documentation says to set CURSOR_API_KEY or pass an API key; it accepts user API keys and service-account API keys for local and cloud runs, but not Team Admin API keys. The example that explicitly selects Composer 2.5 uses the local runtime. The documentation retrieved here does not establish that this model ID is supported through the Cloud Agents API, so verify that route’s supported models before relying on it. (cursor.com)

SDK runs follow the same pricing, request pools, and Privacy Mode rules as runs from the IDE and Cloud Agents. Composer 2.5 draws from the weekly Cursor-model pool included with paid plans; legacy plans use the monthly Cursor Models pool. Beyond the applicable included pool, Cursor lists $0.50 per million input tokens and $2.50 per million output tokens for standard, or $3 per million input tokens and $15 per million output tokens for Fast. Fast is the default in the product. These are token rates for additional usage, not evidence of a separately available raw-inference endpoint. Check Cursor’s current plan and usage documentation for the terms that apply to your account. (cursor.com)

Bottom line

Composer 2.5 is Cursor’s documented agentic coding model, announced May 18, 2026, with a listed 200k context window. For programmatic use, the documented path is running Cursor agents through the SDK or Cloud Agents API—not calling a standalone Composer completion endpoint. The Python SDK documents selecting composer-2.5 for a local agent run; confirm model-specific support before assuming the Cloud Agents route does the same. (cursor.com)

Sources

Composer 2.5 | Cursor Docs

Composer 2.5 | Cursor Changelog

Cursor Python SDK | Cursor Documentation

Cursor APIs Overview | Cursor Docs

Introducing Composer 2.5 | Cursor

Cursor admits its new coding model was built on top of Moonshot AI’s Kimi | TechCrunch