# Cursor Composer 1: What It Is, What It Costs Now, and What Replaced It

> Composer 1 was Cursor's first in-house coding model; it now bills from prepaid credit, and Cursor recommends the cheaper Composer 2.5 instead.

By Pat Simmons · November 17, 2026 · 8 min read

Canonical: https://www.aiformortals.co/blog/cursor-composer-1

Composer 1 is Cursor's first in-house AI coding model. Cursor launched it with Cursor 2.0 on October 29, 2025 and called it "a frontier model that is 4x faster than similarly intelligent models," a speed claim from Cursor's own testing. It runs inside Cursor's coding agent.

You can still pick Composer 1 in Cursor, but it's billed from prepaid credit at $1.25 per million input tokens and $10 per million output tokens, outside the usage your plan includes. Cursor's own docs now say "We strongly recommend using Composer 2.5," which costs $0.50 and $2.50 for the same amounts and counts toward the weekly usage on paid plans.

Last checked November 17, 2026: Composer 1's price and status on Cursor's docs, Cursor's plan prices, and the launch posts for Composer 1.5, 2 and 2.5, all linked below.

## What is Cursor Composer 1?

Composer 1 is the coding model Cursor trained itself, built to do the work inside Cursor's agent: reading and editing files, running terminal commands and searching a whole codebase. Cursor's launch post calls it a "new agent model designed for software engineering intelligence and speed."

Cursor itself is an app for writing software with AI. Its homepage calls it an "AI Coding Agent for Building Ambitious Software," and you can choose which model its agent runs on, including Cursor's own Composer models and models from OpenAI, Anthropic and Google.

Composer 1 is a mixture-of-experts model, a design where only some parts of the model switch on for each piece of text. Cursor trained it with reinforcement learning: the model tried real software tasks in large codebases, using the same search and editing tools the agent uses, and training rewarded the attempts that solved the task well.

To do that, Cursor ran hundreds of thousands of sandboxed coding environments at once in the cloud, on thousands of NVIDIA GPUs. Composer grew out of an earlier test model codenamed Cheetah, and Cursor describes Composer as "a smarter version of this model."

Composer shipped with Cursor 2.0, the same release that let you run up to eight agents in parallel on one prompt, each in its own isolated copy of your code. Its model ID in Cursor's docs is composer-1, and it has a 200,000-token context window, the amount of text it can keep in view at once.

## How much is Cursor Composer?

Composer 1 costs $1.25 per million input tokens, $0.125 per million cached input tokens and $10 per million output tokens. Tokens are the small chunks of text a model reads and writes, so a long agent session that reads a lot of code and writes a lot back costs more.

_The Pricing section of the Composer 1 page in Cursor's docs, captured November 17, 2026._

That money comes from prepaid credit, a balance you buy in advance. On older Cursor plans that are still running for existing customers, Composer 1 draws from the "Other Models" pool instead.

Composer 2.5, the model Cursor recommends now, is cheaper per token and counts toward the usage paid plans include. Cursor's models and pricing docs say each paid plan includes a weekly pool of usage that covers Composer 2.5 and the Grok models, and the rates below apply once you go past that pool.

ModelInputCached inputOutputIncluded in plan usage?

Composer 1$1.25$0.125$10No, billed from prepaid credit
Composer 2.5$0.50$0.20$2.50Yes, weekly pool on paid plans
Composer 2.5 (Fast)$3$0.50$15Yes, weekly pool on paid plans

Prices per million tokens, from Cursor's docs, checked November 17, 2026.

The plans themselves cost $0 (Free), $8 (Plus), $30 (Premium), $100 (Super) and $200 (Ultra) a month, per Cursor's pricing page. Cursor's docs list Composer 2.5 on Plus, at fixed settings with Fast mode off, but the pricing page marks Plus "Only Grok models are available on this plan," so check the model list in the app before buying Plus for Composer.

For what else each tier includes, see what each Cursor plan costs.

## Can you still use Composer 1 in Cursor?

Yes. Cursor's docs still list Composer 1 with its model ID, its tools and its price, and say it "has access to all agent tools when used with Cursor," from editing files and running shell commands to controlling a browser.

_The top of Cursor's Composer 1 docs page, with the note pointing people to Composer 2.5, captured November 17, 2026._

The same page tells you to move on: "We strongly recommend using Composer 2.5. It's smarter and includes more usage on individual plans." The docs don't call Composer 1 retired or give a shutdown date.

Composer 1's output tokens cost four times what Composer 2.5's do at standard speed, and only Composer 2.5 draws on the weekly usage your plan already pays for.

## How good and how fast was Composer 1?

On Cursor's own launch test, Composer 1 scored above the fast and open models of the time and had the longest speed bar on the chart, while GPT-5 and Sonnet 4.5 still scored higher. Every number here is Cursor's own; Pat Simmons hasn't tested Composer on his channel.

_Cursor's launch chart for Composer, October 29, 2025. Composer's speed bar is the longest, and its quality bar sits below "Best Frontier."_

In the launch post's footnote, "Best Frontier" means GPT-5 and Sonnet 4.5, "which both outperform Composer." "Fast Frontier" covers fast models such as Haiku 4.5 and Gemini Flash 2.5, and "Best Open" covers open-weight models such as Qwen Coder and GLM 4.6. The scores come from CursorBench, Cursor's internal test built from real requests by its own engineers and researchers, with hand-picked best answers.

Cursor later published three scores for Composer 1 in its Composer 2 post: 38.0 on CursorBench, 40.0 on Terminal-Bench 2.0 (a test of agents working in a computer's terminal) and 56.9 on SWE-bench Multilingual (coding tasks taken from public code projects).

That 38.0 sits well below Composer's bar on the launch chart. Cursor's post on CursorBench says it keeps updating the test with harder problems, and that scores should only be compared within the same version.

## Composer 1 vs 1.5, 2 and 2.5: what changed

Each new Composer scored higher on Cursor's own tests. Composer 1.5 gave the same base model 20 times more reinforcement learning, Composer 2 switched to a stronger base, and Composer 2.5 trained that base on harder tasks.

VersionReleasedHow Cursor built itCursorBenchTerminal-Bench 2.0SWE-bench MultilingualPrice per million tokens (in / out)

Composer 1Oct 29, 2025Mixture-of-experts model trained with reinforcement learning; base model not disclosed38.040.056.9$1.25 / $10 (today)
Composer 1.5Feb 9, 2026Same pretrained model, 20x more reinforcement learning; thinks before answering44.247.965.9Not on the pages checked
Composer 2Mar 19, 2026First continued pretraining run, on Moonshot's Kimi K2.5, plus reinforcement learning61.361.773.7$0.50 / $2.50 (Fast: $1.50 / $7.50)
Composer 2.5May 18, 2026Same Kimi K2.5 base as Composer 2, 25x more synthetic training tasks63.2 on the newer CursorBench 3.1 (Composer 2: 52.2)69.379.8$0.50 / $2.50 (Fast: $3 / $15)

All scores are Cursor's own. The first three CursorBench scores come from one version of the test; Composer 2.5's comes from the harder version 3.1, so compare it only with Composer 2's 52.2.

Composer 1.5 arrived on February 9, 2026. Cursor built it "by scaling reinforcement learning 20x further on the same pretrained model," and made it a thinking model, one that reasons through a problem before it answers, briefly on easy tasks and longer on hard ones. It can also summarize its own work when it runs out of room, so it keeps going on long tasks. Cursor describes it as "a significantly stronger model than Composer 1."

Composer 2 followed on March 19, 2026, built on what Cursor calls its "first continued pretraining run," which gave reinforcement learning a stronger base to work from. Cursor priced it at $0.50 per million input tokens and $2.50 per million output tokens, with a faster version of the same model at $1.50 and $7.50, and made the fast one the default.

Composer 2.5 came out in May 2026. Cursor says it is "better at sustained work on long-running tasks" and "follows complex instructions more reliably" than Composer 2. In the same post, Cursor says it is training a significantly larger model from scratch with 10x more total compute.

_Cursor's benchmark table from the Composer 2.5 launch post, May 18, 2026, comparing it with Opus 4.7, GPT-5.5 and Composer 2._

Composer 2.5 is the version the Composer 1 docs point you to. For how Cursor compares with other coding agents for different jobs, see Cursor vs Claude Code, by job, and for each new Cursor release, Cursor's latest updates.

## What model is Cursor Composer built on?

Cursor never said what Composer 1 was built on. Composer 2 and 2.5 are built on Moonshot's open-source Kimi K2.5, as Cursor states in its Composer 2 technical report.

The Composer 1 launch post doesn't say whether Cursor trained the model from scratch or started from someone else's open model. Asked on launch day whether Composer was a fine-tune of an existing open-source model, Cursor's Sasha Rush answered, "Our primary focus is on RL post-training," meaning the reinforcement learning that comes after a model's first round of training. He also called rumors that Cheetah, Composer's predecessor, was based on a Grok model "Straight up untrue."

Composer 1.5 used "the same pretrained model" as Composer 1, so its base is undisclosed too. The Kimi K2.5 base only applies from Composer 2 on. For the business side, see who owns Cursor and the other AI coding companies.

## Questions people ask about Cursor Composer

Is Composer 1 free?
No. On current plans Composer 1 is billed from prepaid credit at $1.25 per million input tokens and $10 per million output tokens, and on legacy plans it draws from the Other Models pool. Composer 2.5 is the Composer model that comes with plan usage.

Is Composer 1 still available?
Yes. It's still listed in Cursor's docs with its price, under a note recommending Composer 2.5.

What came with Cursor 2.0 besides Composer?
Cursor 2.0 added a new editor layout built around agents, running up to eight agents in parallel on one prompt, general release of the agent's built-in browser, and sandboxed terminals on macOS that run agent commands without internet access by default.

Is Composer 1 the same as Composer 2?
No. Composer 2 starts from Kimi K2.5 with a new round of pretraining, while Composer 1's base model was never disclosed. Composer 2 scored 61.3 on CursorBench against Composer 1's 38.0 on the same version of the test, and its listed output price, $2.50 per million tokens, is a quarter of Composer 1's current $10.

Can you use Composer outside Cursor?
Cursor built Composer for its own agent, and its docs describe Composer 1's tools "when used with Cursor." Cursor's launch post doesn't mention a public API for it.

## Sources

- Cursor: Composer, building a fast frontier model with RL

- Cursor changelog: Cursor 2.0

- Cursor Docs: Composer 1

- Cursor Docs: Models and pricing

- Cursor: Pricing

- Cursor: CursorBench

- Cursor: Introducing Composer 1.5

- Cursor: Introducing Composer 2

- Cursor: Introducing Composer 2.5

- Cursor: Composer 2 technical report

- Simon Willison: Cursor Composer launch notes
