Pricing and rate limits

Pricing

You pay only for what you use. Meta Model API bills text models per token, image generation per image, Muse Voice Transcribe per minute of audio processed, and Segment Anything Model (SAM) per image or video frame segmented, with no minimums or upfront commitment.

Standard tier

Models: muse-spark-1.3, muse-spark-1.2, muse-spark-1.1.

Standard pricing; your prompts and completions are not used to train Meta models. These versions share the same standard pricing:

Contributor tier

Models: muse-spark-1.3-contributor, muse-spark-1.2-contributor.

Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models. It lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable.

Muse Voice Transcribe

Muse Voice Transcribe is billed by minutes of audio processed.

Streaming and non-streaming transcription are priced the same. ZDR is priced at parity with Standard, and platform free-tier credits apply. The training-eligible discounted tier is not available for Muse Voice Transcribe at launch. Audio time is billed as a separate line item from token usage (rounded down to whole seconds); other endpoints are not billed for audio seconds.

Shared pricing notes

Cached input costs less. When part of your prompt matches a cached prefix, you pay the cached-input rate for those tokens. Check cached_tokens in the response to see how many were served from cache.

Web search grounding costs $2.50 per 1,000 search queries, in addition to the request's token cost. This applies to the web_search tool on text models such as Muse Spark.

There is no long-context premium: you pay the same rate whether your context window is mostly empty or almost full.

Image generation (Muse Image)

Muse Image is billed at a flat $0.01 per generated image. The price is the same regardless of prompt length, reasoning_strength, or the tools the model uses during generation, including its built-in web and image search, which isn't charged separately. A request that returns n images is billed for n images. You're billed only for images the model successfully generates and returns: images that fail to generate, or that are removed by safety filtering before they're returned, aren't counted. Image responses still include a usage object with token counts for reference, but image generation tokens are not included in token-based pricing for other models. See the Image generation guide for the response shape.

Segment Anything Model 3.1

SAM 3.1 segmentation is billed per image and per video frame, not per token. The same prices apply to streaming and non-streaming requests.

Rate limits

Rate limits control usage for your team and are set by pricing tier or model family.

For token-based models, the Standard tier covers muse-spark-1.3, muse-spark-1.2, and muse-spark-1.1; the Contributor tier covers muse-spark-1.3-contributor and muse-spark-1.2-contributor.

Muse Image (muse-image-1.0) has a separate limit of 150 requests per minute. Because it's priced per image rather than per token, it has no tokens-per-minute limit.

SAM 3.1 segmentation has no separate per-model rate limit; its requests count toward your team's tier limit — 3,000 RPM on the Standard tier (see the table above).

Limits apply per team, not per API key. If you use multiple keys in one team, all requests, tokens, images, and audio minutes count toward the relevant shared quota.

Token usage counts only your input and output tokens. Meta injects a small amount of steering context into every prompt (a system prompt and related scaffolding); those injected tokens are not billed and are excluded from the token counts reported in your usage and by the token-counting endpoints.

Retry on 429 responses

If you exceed a request, token, image, audio, or stream-start limit, the API returns a rate-limit response. Requests succeed again once usage drops below the limit.

Retry strategy: Use exponential backoff with jitter. Start with a short delay such as 500 ms, double it after each consecutive rate-limit response, and add random jitter to avoid thundering-herd effects across concurrent clients.

Read rate-limit headers

Every successful token-based response includes headers that report your current limit and remaining quota:

Read these headers and slow down before you hit a rate limit instead of reacting after one.

Submit background responses within limits

Creating background responses (background: true) counts against a separate per-team submission limit in addition to the RPM and TPM limits above. The default is 600 background submissions per minute, per team.

If you exceed that cap, new background submissions return HTTP 429 Too Many Requests with a Retry-After header. Responses already running continue unaffected. Apply the same exponential-backoff-with-jitter strategy described above.

The cap applies only to starting new background responses. Standard RPM and TPM limits still apply to every request.

Rate limits for Muse Voice Transcribe

Muse Voice Transcribe is limited by stream usage rather than tokens. Two dimensions are enforced at once:

Concurrent streams caps how many streaming sessions you hold open at the same moment. Streams per hour limits how many transcription sessions you start.

Both realtime and file transcription use the same budget. A rejected realtime stream closes with 1013. A rejected file-transcription request returns HTTP 429 Too Many Requests.

Next steps

• Muse Voice Transcribe: stream audio or transcribe a file
• Authentication: create and manage API keys
• Get started: make your first API call
• Chat completion: start generating text with the core conversational endpoint