Meta released two models on October 12, Muse Wonder and Muse Wonder Full, per Meta. Muse Wonder is on the Meta Model API. Muse Wonder Full is "available exclusively by subscription in Muse Code," Meta's coding tool. Neither has downloadable weights.
Both models have a 1 million token context window. Muse Wonder costs $4 per million input tokens and $20 per million output tokens, with cached input at $0.20. That matches the list price of Claude Opus 5.5 and sits well under GPT-6 Astra at $10 and $50, per VentureBeat. Wonder Full comes with the Power Usage plan at $50 a month and the Maximum Usage plan at $200. It has no pay-as-you-go price.
On the independent Artificial Analysis index, Wonder Full scores 56 and Wonder 52. Claude Opus 5.5 leads at 58, and GPT-6 Astra and Gemini 4 Argon sit at 53. Meta's own table shows Wonder Full ahead on long coding runs, with 57.5 on SWE-Marathon against 46.3 for Opus 5.5. It trails on Terminal-Bench 4.0 and HealthBench Pro. VentureBeat notes the table mixes third-party runs, Meta-run tests, and vendor-supplied scores.
The subscription carries a data term. Muse Code plans default to what Meta calls Contributor mode, where prompts, code, and outputs may be used to improve future models. Users can turn it off for future submissions. The pay-as-you-go API runs in a mode where submitted content is not used for training. The only route to the stronger model is the one that defaults to sharing your code.
Bottom Line
Meta now has a model within two points of the leader on an independent index, at the leader's price, and it kept the better one inside a subscription. If you evaluate Wonder through the API, you are testing the weaker model. If you try Wonder Full in Muse Code, switch off Contributor mode before you point it at a private repository.