The Meta Muse API: What a Developer Actually Gets, and Which Muse Models It Serves

There is no product literally called the Muse API. The API that serves Meta's Muse models is Meta Model API, it answers at https://api.meta.ai/v1, and the Muse models a developer can call on it are Muse Spark 1.3 (the one Meta currently recommends), Muse Spark 1.2 and Muse Spark 1.1 — the last two of which this site routes, as meta/muse-spark-1.1 and meta/muse-spark-1.2. If you searched for "muse api" wanting a one-line answer, that is it. The rest of this page is the detail behind it: what each model costs per million tokens on Meta's published rate card, what you need in hand before the first call, what the rate limit is actually measured against, and why the numbers on a Muse Code subscription card do not carry over to an API bill.

Read this as a reference page rather than as news. Meta Model API is live and the Muse models on it are callable today; nothing here is an announcement, and no page on this blog has previously answered the programmatic-access question for this model line. Where a figure comes from Meta's own documentation this page says so, and where Meta publishes nothing the page says that instead of filling the space.

What "the Meta Muse API" actually is

Meta's own definition is a single sentence: Model API is "the inference platform for the latest models from Meta including Muse Wonder and Muse Spark." The Muse line is hosted on that platform rather than on an API of its own, which is why a search for a Muse-specific endpoint lands nowhere — the endpoint is Meta's, and Muse is one family of models on it.

The surface is deliberately conventional. Three request formats are documented under what Meta calls Model API surfaces: the Responses API, the Chat Completions API, and the Messages API. The headline integration claim is that Model API is "drop-in compatible with the O​penAI SDK, the A​nthropic SDK, and O​penAI-compatible agent CLIs," with the accompanying line that there is "no custom client to learn." In practice that means a base-URL change and a key, not a rewrite.

The platform serves more than the Spark line. Alongside muse-spark-* it exposes muse-wonder-1.0, muse-image-1.0, muse-voice-transcribe-1.0 and sam-3.1, so a reader building a mixed pipeline can reach image generation, transcription and segmentation on the same credential. One Muse-family member is not callable at all: Muse Glimmer is an open-weight model "built to run on your own hardware," distributed as weights rather than served on the API, and Meta excludes it from the models list endpoint. If your question is "can I call Glimmer," the answer is no — you download it.

Two things people routinely fold into "the Muse API" are separate products and belong on other pages. Muse Code is Meta's coding agent for the terminal and CI — a harness sold as a subscription, shipped as a beta on 2026-08-05 and taken out of beta on 2026-08-31 — and it is not an API, not a model, and not required in order to call the API. This blog's muse-code page covers the harness and meta-muse-code-terminal-coding-agent covers the agent in depth. What follows here is only the programmatic surface.

The models you can call, and Meta's published rates

Six ids matter for text generation, and the rates below are Meta's own, published on its pricing and rate-limits page in dollars per million tokens. They are vendor list prices, not independently audited figures.

• muse-spark-1.3 — Meta calls it "the latest version, tuned for agentic workflows" and marks it "Recommended for new work." 1,048,576-token context. Standard mode: $1.25 input, $4.25 output, $0.15 cached input.

• muse-spark-1.2 — Meta's models page describes it plainly as "the previous version." Same 1,048,576-token context, same Standard rates.

• muse-spark-1.1 — "the original version," and Standard mode only. Same context window, same Standard rates again.

• muse-spark-1.3-contributor and muse-spark-1.2-contributor — the discounted tier: $0.10 input, $0.20 output, $0.002 cached input. There is no Contributor variant of 1.1.

• muse-wonder-1.0 — Meta's most capable Muse model, positioned for long-horizon agentic work: $4.00 input, $20.00 output, $0.20 cached input, Standard mode only, and Meta states plainly that "Contributor mode isn't available for Muse Wonder on Model API."

• Non-text Muse models are priced per unit rather than per token: Muse Image at $0.01 per image, Muse Voice Transcribe at $0.18 per hour, and SAM 3.1 at $2.50 per 1,000 images and $0.20 per 1,000 frames. Web search grounding, if you turn it on, is $2.50 per 1,000 search queries.

Two general statements sit under that table and are worth knowing before you size a workload. The first is "Cached input costs less" — cached reads are roughly a tenth of input on Standard Spark, and a fiftieth on Contributor. The second is "No long-context premium," which matters more here than on most platforms because the Spark line's window is 1,048,576 tokens; a long document does not move you onto a different price.

On the question the brief for this page expected to be a contrast: between meta/muse-spark-1.1 and meta/muse-spark-1.2 there is nothing to choose on window or on rate. Meta publishes one context figure for the line and one rate row covering "muse-spark-1.3 / 1.2 / 1.1." What separates them is the checkpoint — 1.2 is the later one — and, on Meta's platform, the presence of a Contributor variant for 1.2 and not for 1.1. Anyone who tells you the two differ in context length or in price is not reading the same cards.

What you need in hand to make the first call

Everything in this section is from Meta's own documentation. Where a thing Meta does not say would normally go, this page leaves the gap visible rather than inventing a plausible request shape.

Getting access

Access is a Model API account plus a key. Meta's quickstart directs you to the dashboard: "Get a key from the Model API dashboard → API keys → Create API key." The authentication page adds the part people miss: "Copy the key right away. You only see it once." A key is displayed in the form LLM|607358788350350|nx9.....LJY, and Meta never explains what the segments mean. You can revoke a key from the same dashboard — "Deleting one key leaves other team keys untouched" — which is Meta's argument for issuing one key per application rather than sharing one.

The auth scheme

Bearer token, in the Authorization header:

• Header — Authorization: Bearer $MODEL_API_KEY

• Base URL — https://api.meta.ai/v1, with https://api.meta.ai documented for A​nthropic-format harnesses

• Environment variable — MODEL_API_KEY, which Meta instructs you to set rather than hard-code, alongside the usual "never commit keys to version control" advice

• Client caveat — Meta states outright that "The O​penAI SDK doesn't auto-read MODEL_API_KEY," so you either pass api_key explicitly or export OPENAI_API_KEY with your Meta key

Endpoints named across Meta's reference include POST /v1/chat/completions and POST /v1/responses for text, /v1/images/generations and /v1/images/edits for Muse Image, POST /v1/asr/transcribe plus a realtime websocket at wss://api.meta.ai/v1/asr/realtime for transcription, and POST /v1/files for uploads. There is no Meta-published SDK: the SDKs page lists the O​penAI and A​nthropic clients, and Meta's position is compatibility rather than a first-party library. If you were waiting for a meta package to install, there is not one.

Rate limits, and what they are measured against

This is the question the "muse code api key" searcher is usually really asking, and Meta answers it more precisely than most vendors do. The published limits are per team, not per key: "Keys on the same team share rate limits." Two engineers minting two keys do not double the ceiling — they split it.

• muse-spark-1.3, muse-spark-1.2, muse-spark-1.1 — 3,000 requests per minute and 4,000,000 tokens per minute

• The Contributor variants — 150 requests per minute and 3,000,000 tokens per minute

• muse-wonder-1.0 — 3,000 RPM and 4,000,000 TPM; muse-image-1.0 — 150 RPM with no token cap

• Background responses — a separate ceiling of 600 submissions per minute per team

• On breach — HTTP 429 with a Retry-After header, and responses carrying x-ratelimit-limit-tokens, x-ratelimit-remaining-tokens, x-ratelimit-limit-requests and x-ratelimit-remaining-requests

A Muse Code API key and a Model API key are the same kind of credential. Muse Code's first run asks you to pick browser sign-in or to paste an API key — "For scripts and CI, use an API key" — and its documented lookup order is META_API_KEY first, then a stored key, then a stored browser session, with an API key always outranking a browser sign-in.

The billing shape: metered tokens, not a seat

This is the section readers get wrong, so it is worth being blunt about it. A Muse Code subscription is a flat monthly seat with a rate card; Model API is metered consumption on an account. Meta's auth page states the model directly — "Meta Model API billing is usage-based: you're billed for the tokens your requests consume" — with input and output counted separately, output typically costing more, and rates varying by model. Money moves when your balance reaches your payment threshold, and any remainder is charged automatically on the first of the month. Meta adds that there are "no minimums or upfront commitment."

One asymmetry between the two products is easy to trip over. Meta states that Muse Code "subscriptions default to Contributor mode, and you can turn it on or off in settings." On the API the default is the opposite: you get Standard pricing and Standard terms unless you deliberately address a -contributor model id. So a reader who has been running Muse Code under Contributor terms all along and then swaps in a key at muse-spark-1.2 will land on the Standard rate card at roughly twelve times the input price, without having changed anything except the client.

What Meta does not state is whether a subscription entitles you to any API usage. Nothing on the authentication page, the pricing page or the product page says a plan includes Model API tokens or credits; the API page treats a key purely as an authentication method, and the subscription pages are their own subject. Do not assume the two are fungible, and do not carry the subscription's monthly figure onto a token budget — they are answering different questions.

On free: there is no free tier documented for Model API, no trial credit, and no free allowance. The product page for the coding agent lists no free plan either. Contributor mode is discounted, not free — Meta's framing is discounted pricing "in exchange for permission to use your content to train future Meta models," covering the prompts, code, files and context you send plus the responses you get back. That is a trade, not a giveaway. If a page tells you the Meta Muse API has a free tier, it is not reading Meta's pricing page.

What this property routes

We serve two models from this line, and the availability claim on this page is deliberately narrow. Both are live in the catalogue today: meta/muse-spark-1.1 and meta/muse-spark-1.2. Both carry the same 1,048,576-token context window Meta publishes and 1.2 is the later checkpoint. Our published rates for them are Meta's list price passed through at 0% markup — $1.25 per million input tokens, $4.25 per million output tokens and $0.15 per million cached reads — which is the whole point of a pass-through arrangement: a vendor rate change is live here the same day, with no repricing on our side.

What that buys a developer is not a discount; it is one fewer integration. Muse Spark sits behind a single endpoint covering 200+ models, so adding it to a pipeline is not a second contract, a second key or a second billing relationship, and a request that fails against one provider fails over to another automatically rather than surfacing a 429 to your users. That matters more than usual with an unproven model swap: it is a way to put meta/muse-spark-1.2 on a real workload before committing the path to it.

Two things are not routable here and this page will not soften that. meta/muse-spark-1.3 is not in our catalogue — if 1.3 is what you need, call Meta directly. And meta/muse-code is not a model id at all, on Meta's platform or on ours, because Muse Code is a harness rather than a model; it cannot be a routing target. The Contributor-tier checkpoints are not routed here either.

What Meta does not document

Several things a developer would reasonably expect are simply absent from Meta's API documentation, and the honest version of this page says so rather than sketching them in.

• Organizations and projects — there is no org or project concept on the authentication page, no hierarchy, and no way to scope a key to one. The only grouping Meta describes is a team.

• Key scopes — no permission sets, no per-resource restrictions, no read-only keys. Every key is a full key, which is why Meta's own advice is one key per application.

• Key lifetime and rotation — revocation is documented (delete it in the dashboard); rotation is only advice ("revoke and rotate immediately"), with no expiry, TTL or rollover procedure anywhere.

• A first-party SDK — none published; the compatibility claim is the integration story, and Meta does not say there will be one.

• Free usage — no free tier, no credits, no trial, as above.

• An API id for Muse Glimmer — the models page lists no id and Meta excludes it from the models endpoint, because it ships as weights.

If your reason for searching this was one of those gaps, the answer is that the information is not published, and a worse page would have guessed at a key-scoping scheme or a rotation endpoint and left you to discover the guess at 2am. On what is published, the picture is complete enough to build against: a conventional base URL at https://api.meta.ai/v1, a Bearer key you create yourself, drop-in compatibility with the O​penAI and A​nthropic clients, documented per-team rate limits, and a rate card in which the two Muse Spark checkpoints this site serves are identical on both window and price.

This blog's muse-code page covers the harness in depth; this page is only the programmatic surface.

Gaps like these are also where failure modes show up: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.