What Is Meta Muse? Model API and Muse Code: Models, Pricing, Contributor Mode and Setup

Meta Muse lets you use Meta's AI models from your own app or from a coding agent in the terminal. With the Meta Model API (dev.meta.ai) you call the models on a pay-as-you-go basis; because the endpoint is OpenAI- and Anthropic-compatible, existing SDKs and agents like Claude Code connect by changing only the base URL and key. Muse Code is a terminal coding agent running on the same models; you use it pay-as-you-go or with a subscription starting at $5/month. There is no free tier. The most important caveat is Contributor mode: in exchange for a lower price or more usage, the content you send is used to train Meta's models; it is on by default in Muse Code subscriptions and should not be used with client or personal data.

What it does: API and terminal coding agent for Meta models
Where you use it: API (your own app), Muse Code (terminal), agents like Claude Code and Codex
Free: No free tier in the official documentation
Contributor mode: Cheaper, but content is used for training
Age: 18
Turkey: Not clear in the official documentation
Price: Pay-as-you-go; Muse Code subscription $5–200/month

What is Meta Muse for?

When a developer wants to add AI to an app, they either use an API (a gateway for sending requests from one program to another) or run a ready-made coding agent. Meta offers both: the Model API opens Meta's text, image, speech and image-segmentation models with a single key; Muse Code is a ready-made agent that plans, edits files and runs commands in your project folder.

Connect a Meta model to your appWhat you get: Text, code, tool calls, structured JSONNeeds a connection Coding agent in the terminalWhat you get: Changed files, commands run, session logNeeds a separate setup Run Claude Code with a Meta modelWhat you get: A session in the Claude Code interface running on a Meta modelNeeds a connection

Who it's for: developers connecting a model to their app, people who want to try Claude Code or Codex with a different model, and anyone looking for a flat-rate monthly coding agent in the terminal. It is not a chat app for people who don't write code.

Two products, same models

	Model API	Muse Code
What	An endpoint you call from your own app	Ready-made terminal coding agent
How you pay	Pay-as-you-go (tokens, images, audio hours)	Pay-as-you-go or monthly subscription
Key	API key you create in the dashboard	Browser sign-in or API key
Contributor mode	Chosen via the model ID	On by default in subscriptions

Meta Model API: key, endpoint and first request

Create an account: "Start building" on dev.meta.ai; with email, phone, or a Facebook, Google or GitHub account. You must be 18 and located in a supported country.
Add a payment method: On the Billing page. Charges accrue as you use; they are billed to your card when a threshold is reached or on the 1st of the month.
Create a key: "Create API key" on the API keys page; keep the key like a password and don't embed it in code.

Endpoints: https://api.meta.ai/v1 for OpenAI-compatible clients (Responses and Chat Completions), https://api.meta.ai for Anthropic-compatible clients (Messages). Meta doesn't require a separate SDK; in clients such as the OpenAI SDK, Anthropic SDK, LangChain or LiteLLM, changing the base URL and key is enough.

Which models are available? (example from the docs)

› curl -X GET "https://api.meta.ai/v1/models" \
  -H "Authorization: Bearer $MODEL_API_KEY"

What you get: the list of models available to your team, newest first. Then a client.responses.create(model="muse-spark-1.3", input="…") call with the OpenAI SDK returns a text response (example from the SDKs and libraries page).

Environment variables (configuration from Meta's docs)

› export ANTHROPIC_BASE_URL="https://api.meta.ai"
export ANTHROPIC_AUTH_TOKEN="$MODEL_API_KEY"
export ANTHROPIC_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_OPUS_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_SONNET_MODEL="muse-spark-1.3"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="muse-spark-1.3"
export CLAUDE_CODE_SUBAGENT_MODEL="muse-spark-1.3"
export ENABLE_TOOL_SEARCH="true"

What you get: When you run the claude command with these variables, Claude Code works with the same interface but on Meta's models; usage is charged pay-as-you-go to your Meta account. The docs ask you to use ANTHROPIC_AUTH_TOKEN instead of ANTHROPIC_API_KEY, because Meta expects the key as a Bearer token. All three aliases (opus, sonnet, haiku) must point to a Meta model; otherwise plan mode and subagents look for a model Meta doesn't serve.

OpenAI-compatible agents connect to https://api.meta.ai/v1. Meta's quickstart page provides a ready opencode.json block for OpenCode and a setup instruction to paste into the agent; the separate configuration for Codex is in the "coding agents" doc.

Rate limits apply per team, not per key. Text models: 3,000 requests per minute; Contributor variants: 150. The web search tool costs an extra $2.50 per 1,000 queries.

Muse Code: installation and use

Muse Code is Meta's coding agent for the terminal and CI (automated test/build pipelines). You run it in a project folder; it plans the task, edits files and runs commands. Approval modes and an OS-level sandbox are on from the first run. According to the docs it runs on macOS, Linux and Windows; the product page says "macOS and Windows".

Install› curl -fsSL https://dev.meta.ai/install.sh | sh

Install› irm https://dev.meta.ai/install.ps1 | iex

On Windows, agent commands run through PowerShell; the sandbox may ask for administrator approval on first launch; voice input and cross-session messaging are not available.

These install commands download a script from the internet and run it directly; if you like, open the URL in a browser first and read the script. You can confirm the installation with muse --version.

Start (in the project folder)

› cd /path/to/your/project
muse

On first launch Muse Code asks whether you trust the folder and asks for browser sign-in or an API key. Then you describe the task in plain language.

Debugging"The tests are red. First find which test fails and why, then make the smallest fix and run the tests again."What you get: according to the docs, the agent plans, edits the file and runs the command; every shell command goes through the approval mode and sandbox rules. At the end, the changed files and the commands run are kept in the session log.

One-off run from CI or a script› META_API_KEY="$MUSE_KEY" muse exec "Run the test suite and summarize failures."

What you get: in a non-interactive run the prompt runs to completion and returns a summary (example from the docs). To drive it from your own program, muse serve and npm install @muse-code/sdk are documented.

The Muse Code docs cover only the command line and SDK installation; an installation path for a desktop app or VS Code extension is not clear in the official documentation.

Which model? Differences

Meta's docs list six model lines. For text and code work the choice is between two models: Muse Wonder is the most capable, Muse Spark is for speed, volume and cost. Both use the same endpoints in the same way; switching from one to the other is just a change of model ID.

Muse Spark 1.3Muse WonderMuse Wonder Full

Volume and costLong, complex work

API + Muse CodeAPI + Muse CodeMuse Code Power and Maximum only

Model	What it's for (Meta's description)	Where	Pay-as-you-go price
Muse Wonder (`muse-wonder-1.0`)	Complex software work, long-running agent work, computer use; in preview	API (Standard only) + Muse Code	Per 1M tokens: input $4, cached input $0.20, output $20
Muse Wonder Full	More capable variant of Muse Wonder	Muse Code Power and Maximum subscriptions only	Included in subscription
Muse Spark 1.3 (`muse-spark-1.3`)	Agent and coding work, high volume; recommended version for new work. Audio input not fully supported	API + Muse Code	Per 1M tokens: input $1.25, cached $0.15, output $4.25. Contributor: $0.10 / $0.002 / $0.20
Muse Spark 1.2	Previous version; recommended for audio input. Muse Code starts with it on first launch	API + Muse Code	Same as Muse Spark (including Contributor)
Muse Spark 1.1	First version	API only (Standard)	Same as Muse Spark
Muse Image (`muse-image-1.0`)	Text-to-image generation and editing; web and image search included in the price	API + Muse Code plans	$0.01 per image
Muse Voice Transcribe	Transcribes live or recorded audio; 25 languages including Turkish, speaker separation. Does not generate speech	API	$0.18 per processed audio hour
SAM 3.1	Finds and masks a described object such as a "yellow school bus" in images and video	API	$2.50 per 1,000 images; $0.20 per 1,000 video frames
Muse Spark Open	Open-weight version: 816 billion total, 42 billion active parameters. Runs on your own GPUs or a hosting provider	Download (Hugging Face) or providers like OpenRouter, Groq, Together AI	Free open weights; Muse Model License (below)
Muse Glimmer	Open-weight model distilled from Muse Spark that runs on your own hardware	Download	Free open weights; Apache-2.0

Multi-file code changes, agent work lasting hours: Muse Wonder.
Classification, short tool loops, subagent tasks, high volume: Muse Spark 1.3.
Requests with audio: Muse Spark 1.2, or Muse Voice Transcribe for transcription only.
Keeping data on your own servers: Muse Spark Open; BF16 needs about 2 TB of GPU memory (the NVFP4 version runs on a single Blackwell node with 4 GPUs).

Allowed

Muse Spark Open: free to use, modify, fine-tune, distill, distribute and sell, at no cost
Muse Spark Open: powering a feature of your own product, internal use and research need no additional agreement
Muse Spark Open: Meta claims no ownership of outputs
Muse Glimmer: Apache-2.0

Not allowed

Every product or service using Muse Spark Open must display "Built with Muse Spark Open" (in the interface, documentation, website or settings menu)
A company offering the model to others as a service through an interface such as an API, with gross revenue over $20 million in the past 12 months, must notify Meta first and sign a separate written agreement
The Acceptable Use Policy applies; not intended for under-18s

Source: Muse Model License (Muse Spark Open version, October 15, 2026) and the Muse Glimmer docs. This is not legal advice.

Muse Wonder and Muse Spark accept text, image, video, audio and PDF input; the context window is 1,048,576 tokens. The amount of reasoning is set with reasoning_effort from "minimal" to "max".

Source: dev.meta.ai/docs/models and /docs/pricing-rate-limits (October 29, 2026); Meta changes versions often. Muse Wonder and Wonder Full are currently unavailable in South Korea, and Muse Image in India.

Pricing: pay-as-you-go and Muse Code subscriptions

Model API is pay-as-you-go only: text models per token, image generation per image, transcription per audio hour, segmentation per image or frame. The lowest text price in Standard mode is $1.25 per 1M input tokens; per-model prices are in the table in the "Which model?" section. There is no surcharge for long context; cached input is cheaper. No free tier appears in the official documentation.

Muse Code can be paid two ways: pay-as-you-go with an API key, or a flat monthly subscription. The subscription applies only in the Muse Code command line when signed in with a Meta account; other keys you create on the same account are billed pay-as-you-go.

Everyday Usage$5 /month

Text, code and image models
10-50 prompts per 5 hours (including image and video uploads)
Voice mode
Web search
Most capable variant (Full)

High Usage$15 /month

Everything in Everyday
5x the usage of Everyday
More multimodal input
Most capable variant (Full)

Power Usage$50 /month

Everything in High
Most capable variant (Full)
20x the usage of Everyday
Early access to new features
Larger file uploads

Maximum Usage$200 /month

Everything in Power
80x the usage of Everyday

You buy and manage the subscription in Meta Accounts Center (Accounts Center › Subscriptions › Muse Code); typing /upgrade in a session opens it.
Upgrades take effect immediately with a prorated charge; downgrades apply from the next billing period.
Cancellation takes effect at the end of the period; cancel at least 24 hours before renewal. No refunds on cancellation (unless required by law).
If you hit the usage limit, you move to a higher plan or wait for the limit to reset.

Prices in US dollars, excluding tax; from the Muse Code product page and subscription docs (October 29, 2026). The docs say subscription benefits and prices may vary by region and are shown before you subscribe. No local-currency (Turkish lira) price is published, and card payment from Turkey is not clear in the official documentation.

Official pricing page last checked on November 3, 2026.

Can you use it from Turkey? Turkish, age and account

Access from TurkeyUnclear	Turkey is not on the list of prohibited regions in the geographic usage policy, but the service is only available in countries where Meta has opened access. The help page says "most countries where Meta AI is available and payments are supported". The sign-up screen tells you whether your country is accepted.
Replies in TurkishUnclear	The help page says the models are multilingual; there is no specific statement about Turkish.
Turkish interfaceNo	The Model API site is English only (help page).
Turkish transcriptionYes	Turkish is among the 25 languages Muse Voice Transcribe supports.
Age18+	At least 18 to sign up and use it; for subscriptions, 18 or the age of majority in your country.
Local currency and card paymentUnclear	Prices are in US dollars; a Turkish lira price and acceptance of Turkish cards are not stated in the official documentation.

Meta does not claim copyright over generated content; under the terms of use, the rights to inputs and outputs are yours. If you publish your own product, your end users must also be over 18 and you must have a privacy policy (help page).

Contributor mode: is your data used for training?

Meta offers two modes. In Standard mode you pay the standard price and the content you send is not used to train Meta's models. In Contributor mode you get a lower price (API) or more usage (Muse Code subscription); in return, the prompts, code, files, context you send and the responses you receive may be used to train Meta's future models.

Meta's help page asks that no sensitive, confidential or personal information be sent in Contributor mode; any use that processes personal data should happen in Standard mode. Don't use Contributor mode with client data, employee information or confidential code; under KVKK (Turkey's Personal Data Protection Law) you would also be exposing that data to training abroad. This is not legal advice.

Model APIYou choose	Contributor applies only when you call variants whose model ID ends in `-contributor`. The normal ID is Standard. Muse Wonder has no Contributor option in the API.
Muse Code subscriptionOn by default	On by default in subscriptions. To turn it off, type `/settings` in a session, select the "Contributor mode" row under Model and press Enter.
Muse Code, pay-as-you-goNo setting	The Contributor row appears only with an active subscription; otherwise it follows the model ID.
Zero data retention (ZDR) accountOff	Teams with ZDR cannot use Contributor variants.

The Muse Code setting is stored per computer: if you use it on several computers, turn it off on each one. On a new computer, the first interactive session turns the mode on and tells you so. For script and CI runs, pass the --contributor-mode off flag or the MUSE_CONTRIBUTOR_MODE=0 variable every time.

Off for a single run› muse exec --contributor-mode off "Summarize the changes on this branch"

Frequently asked questions

What is Meta Muse?

Meta's Muse models for developers and two products: the Meta Model API, which you call from your own app, and Muse Code, a terminal coding agent running on the same models.

What is the difference between Muse Spark and Muse Wonder?

Muse Wonder is the most capable model, for long and complex code and agent work. Muse Spark is cheaper and faster, for high-volume, cost-sensitive work. Both use the same endpoints.

Is the Meta Model API free?

No, the official documentation lists no free tier. You pay as you go; Muse Spark in Standard mode costs $1.25 per 1M input tokens and $4.25 per 1M output tokens.

What is Contributor mode?

A mode where, in exchange for a lower price or more usage, you allow the content you send to be used to train Meta's models. Meta asks that no personal or confidential data be sent in this mode.

How do I turn off Contributor mode in Muse Code?

Type /settings in a session, select the Contributor mode row under Model and press Enter. The setting is stored per computer; for a single run, use the --contributor-mode off flag.

How much does a Muse Code subscription cost?

Everyday $5, High $15, Power $50, Maximum $200 per month. Prices and benefits may vary by region; no refunds on cancellation.

Is the Meta Model API available in Turkey?

Not clear in the official documentation. Turkey is not on the prohibited regions list, but the service is only available in countries where Meta has opened access. If the sign-up screen does not accept your country, the service is not open in your region.

Can I run Claude Code with Meta models?

Yes. Meta's docs explain setting ANTHROPIC_BASE_URL to https://api.meta.ai and ANTHROPIC_AUTH_TOKEN to your Meta API key, and pointing the model variables to Meta models. Usage is charged to your Meta account.

Related guides

Sources

Not clear in the official sources: Whether Turkey is among the countries with access · Turkish lira pricing and payment with a Turkish card · Quality of replies in Turkish · Installation path for a Muse Code desktop app or VS Code extension (the Muse Code docs cover only the command line and SDK). Each is noted in the relevant section.