Published: 13 October 2026
Meta released Muse Wonder and Muse Wonder Full on 12 October 2026, the first models it aims at the top of the market. The API version costs exactly what Claude Opus 5.5 costs, $4 per million input tokens and $20 per million output tokens, yet scores lower on the Artificial Analysis index. Wonder Full, the stronger of the two, is only available inside a Muse Code subscription.
Wonder Full wins on legal and long coding work, Opus 5.5 on terminal and health tasks
Both models come from Meta Superintelligence Labs, the unit Mark Zuckerberg set up in 2025. They read text, images, audio and video, answer in text, and take about 1 million tokens of context (the amount of input they can read at once). Meta sells them on two features. Active Memory carries context from one conversation to the next. Active Focus keeps an agent working through blockers and makes it ask before a consequential step.
The scores split by type of work. Meta leads where a task runs for a long time or follows legal procedure, and trails where the task is short and exact.
| Measured (max reasoning effort) | Wonder Full / Wonder | Claude Opus 5.5 |
|---|---|---|
| Artificial Analysis Intelligence Index (broad average) | 56 / 52 | 58 |
| Harvey LAB-AA (legal agent tasks) | 13.1 / 10.0 | 4.2 |
| SWE-Marathon (long coding tasks) | 57.5 / 53.1 | 46.3 |
| DeepSWE 1.1 (software engineering) | 79.0 / 78.8 | 74.2 |
| Terminal-Bench 4.0-AA (command-line tasks) | 57.1 / 52.0 | 59.6 |
| HealthBench Pro (medical questions) | 62.3 / 60.5 | 65.6 |
The index row comes from Artificial Analysis, an independent benchmarking firm. The other rows are from Meta’s comparison table as reported by VentureBeat. Meta’s model card says rival runs were best effort, and that the table mixes third-party, provider-reported and internal results. It also says SWE-Marathon is not a controlled comparison, because some rivals ran inside their own coding tools. Read the coding leads as a reason to test, not as a settled result.
Opus pricing to the cent, a lower index score and a cheaper bill per task
Meta’s pricing page lists muse-wonder-1.0 at $4 for input and $20 for output per million tokens, with no surcharge for long prompts. That is the Opus 5.5 list price exactly, for a model that scores 52 on the index against 58 for Opus.
The bill can still come out lower. Artificial Analysis puts the average cost of one index task at $4.52 for Wonder, $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra, according to VentureBeat. At equal token prices, that gap suggests Wonder spends fewer tokens on each task.
- Model ID: muse-wonder-1.0 on the Meta Model API, Standard mode only
- Price: $4 input, $20 output and $0.20 cached input per million tokens; web search adds $2.50 per 1,000 queries
- Context: 1,048,576 tokens; text, image, video, audio and PDF in, text out
- Rate limits: 3,000 requests and 4 million tokens per minute, per team
- Weights: none, unlike Meta’s Apache 2.0 model Muse Glimmer
- Data: Meta does not train on content sent in Standard mode
Wonder Full only runs in Muse Code, where Meta trains on your input by default
Wonder Full has no API price and no date for one. You reach it through Muse Code, Meta’s coding tool, on the $50 or $200 monthly plans. Meta states the usage allowance of those plans only as a multiple of its cheapest plan.
The catch sits in Meta’s developer terms. Coding subscriptions run in Contributor mode by default, whichever model you choose, and in that mode Meta may use your prompts, code and outputs to train future models. You can switch it off in your account settings, but only for new input. The same terms forbid sending confidential or personal information to Contributor mode at all.
For European teams that settles it for now. Client code and files containing personal data cannot go into Muse Code until someone changes that setting, and earlier input stays under the training terms. Meta Platforms Ireland is the contracting party for users in Europe, but Meta names no EU data region, and its model documentation only says availability can vary by country.
Trial the API on legal drafting and long coding runs, and keep client code out of Muse Code
Most teams already on Claude Opus 5.5 or GPT-6 Astra have no reason to switch. At the same token price, Wonder scores lower on the broad index and on terminal work, and the full safety report Meta promises has not appeared yet.
Two groups should run a trial now. Legal and insurance teams building document agents come first, because Harvey’s legal benchmark shows Meta’s widest lead. Engineering teams whose coding agents run for hours come second, since SWE-Marathon and DeepSWE point Meta’s way, even if part of that comparison was not controlled.
Start with the API in Standard mode, which keeps your data out of training. Rerun one finished task that took your current agent several hours. Compare the cost per completed task rather than per token, and check that Active Memory does not carry details from one client file into the next. Leave Wonder Full aside until Meta offers it outside a subscription that trains on your input by default.
For more information, visit the official announcement of Muse Wonder on the Meta blog.
