Muse Wonder Preview on the Meta Model API: $4/$20 per 1M Tokens, Wonder Full in Muse Code Only

Our verdict: test it, do not migrate to it. Muse Wonder is a preview at a frontier price. At $4 input and $20 output per 1M tokens it costs 3.2x Muse Spark 1.3 on input and about 4.7x on output, so it is not a cost-neutral upgrade. Run it on the tasks Muse Spark gets wrong, and keep production where it is until there is an outside evaluation and the safety report Meta says is coming. Our directory rates Muse Wonder Conditional, and we have not tested it ourselves.

What happened

Meta announced the Muse Wonder series on 12 October: Muse Wonder and Muse Wonder Full, which it calls "our most capable models". The announcement says Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full is available to Muse Code subscribers.

From Meta's developer docs:

• Model id: muse-wonder-1.0
• Context window: 1,048,576 tokens; text, image, video, audio and PDF in, text out
• Status: "Muse Wonder is in preview."
• Reasoning effort: "minimal" to "max"
• Availability gap: neither model is currently available in South Korea

Pricing per 1M tokens, from Meta's pricing page:

Muse Wonder is priced in Standard mode only: Contributor mode, the cheaper tier that lets Meta train on your content, "isn't available for Muse Wonder on Model API". The rate limit is 3,000 requests and 4,000,000 tokens per minute per team, the same as Muse Spark 1.3.

Why it matters

Wonder Full is not on the API. Meta's models page: "Muse Wonder Full, the higher-capability variant, is available only in Muse Code with a subscription, so setting reasoning effort to "max" on Muse Wonder doesn't select it." It is limited to the Power Usage and Maximum Usage plans. We did not read those plans' prices. API buyers get the lesser variant, whatever effort setting they pick.

The subscription path has a data default. The same page says Muse Code subscribers can use Contributor mode, "which is on by default", giving more usage in exchange for letting Meta train future models on their content, and that it can be turned off in settings.

The evidence is all Meta's. Meta reports results on Harvey's Legal Agent Benchmark, FocusBench and SWE-Marathon, among others. Meta reports every one of these results itself, and we read no independent evaluation. Meta says "a full report with additional safety evaluations will be available soon"; it was not out when we read the announcement.

What changes for you

• On Muse Spark via the API: nothing by default. Route only the tasks Spark fails to muse-wonder-1.0 and budget them at $4/$20 per 1M tokens.
• Running production: stay on a non-preview model until independent results or the safety report appear.
• Considering Muse Code for Wonder Full: you need Power Usage or Maximum Usage, and if your code must not feed Meta's training, switch Contributor mode off before your first session.