Meta Muse Wonder API pricing, per agent task

Muse Wonder API pricing is $4.00 per 1M input tokens, $0.20 per 1M cached input tokens and $20.00 per 1M output tokens for `muse-wonder-1.0` on the Meta Model API (Meta's pricing page, read 2026-10-13). Web search grounding adds a fee per query. There is no cheaper Contributor mode for this model.

A price per token does not tell you what a task costs. An agent on Muse Wonder spends on two meters: tokens with Meta, and every data call it makes along the way. This post puts Meta's rates in one table, works through the arithmetic for one task with the token counts stated as an assumption, adds the data-call line, and lists the limits that cap your spend rate. Model facts come from Meta's models page, read 2026-10-13.

Muse Wonder API pricing at a glance

Meta bills text models per token on pay as you go, with no minimum and no upfront commitment. Muse Wonder has one price mode, Standard. Muse Spark sits on the same page for comparison, since Meta tells you to move work between the two by changing the model id.

Four notes from the same page change the math:

• Cached input is billed when part of your prompt matches a cached prefix. The response's `cached_tokens` field says how many tokens that covered.
• No long-context premium. The rate is the same with a nearly empty or a nearly full context window, and Muse Wonder's window is 1,048,576 tokens.
• Steering tokens are free. Meta adds a small system prompt to every request and does not bill those tokens or count them in your usage.
• Contributor mode is not available for Muse Wonder on the API. Only Muse Spark 1.3 and 1.2 have it.

Muse Wonder is in preview, per Meta's models page, so these rates can change. Muse Wonder Full is not on the API at any price. It runs only in Muse Code on the Power Usage and Maximum Usage plans, and those plans are priced on Meta's subscriptions page, not per token.

Muse Wonder API pricing per agent task: the arithmetic

Add up three token lines and the tool calls. The token counts below are an example we chose, not a measurement: a research task that reads 200,000 input tokens, of which 150,000 come from a cached prefix, writes 20,000 output tokens and runs 5 web searches through Meta's grounding.

cost = uncached_input x 4.00 / 1,000,000
     + cached_input   x 0.20 / 1,000,000
     + output         x 20.00 / 1,000,000
     + web_searches   x 2.50 / 1,000

Two things move this number most. Output is five times the input rate, so a task that writes long answers or long code costs more than one that reads a lot. Caching cuts the input rate to 5 percent of the uncached rate, so an agent that keeps its system prompt and tool list stable from call to call pays far less for the same context.

Your real counts come back in each response's usage object. Run one task, read the counts, and plug them in before you budget a thousand.

The second line on the bill: data calls

Meta's built-in `web_search` tool costs $2.50 per 1,000 search queries on top of the request's tokens (Meta's pricing page, read 2026-10-13). It applies to text models such as Muse Wonder and Muse Spark. When the task needs something else, the full text of a page, news results as rows or Google positions, your agent calls a data tool, and whoever serves it bills that call.

With looot, the agent calls those jobs from one prepaid balance and reads each price before it runs: `job:web.search`, `job:web.scrape.markdown`, `job:news.search` and `job:google.serp.organic`. The table reads the cheapest live price for each job from the looot catalog when the page loads.

For the example task, add each price above times the number of calls. Searching and reading pages is covered job by job on the web search API page and the URL to markdown page. The same two-meter view for any agent, not only Muse Wonder, is in what an AI agent's data calls cost, and the count-times-price math for lookups is in what 1,000 lookups cost.

This prompt prices the data side of one task without spending anything:

Rate limits and caching on the Muse Wonder API

Rate limits cap how fast you can spend. For `muse-wonder-1.0`, Meta sets 3,000 requests per minute and 4,000,000 tokens per minute. The limits apply per team, not per API key, so every key in a team draws on the same budget, and each model has its own limits.

When you pass a limit, the API rejects the request with HTTP 429. Meta's advice is exponential backoff with jitter, starting near 500 ms, and waiting at least as long as any `Retry-After` header. Each response also carries `x-ratelimit-remaining-tokens` and `x-ratelimit-remaining-requests`, so an agent can slow down before it hits the wall.

Caching is the cheapest lever. Keep the stable parts of the prompt, the system prompt, tool definitions and long reference documents, at the start and unchanged, so later calls match the cached prefix and pay the cached rate.

What this post cannot tell you

• A measured task cost. We did not run a Muse Wonder task for this post. The example uses token counts we chose.
• Muse Code plan prices. Muse Wonder Full is subscription only, and we quote no plan price here. Read Meta's subscriptions page.
• Prices after the preview. `muse-wonder-1.0` is in preview, so Meta can change the rates.
• Where you can buy it. Meta says model availability can vary by country, under its Geographic Use Policy.
• Scheduled spend. Running the same task on a timer needs scheduled runs, not available yet in looot.

Last checked 2026-10-13. Meta prices and limits were read that day from the Meta pages linked above. looot prices in the table are read from the live catalog when the page loads. More on the model itself is in what is Muse Wonder.