Muse Wonder is Meta’s new family of AI models for writing code and handling jobs that take many steps and several software tools. Meta released two versions, Wonder and Wonder Full, on October 12 in its Muse Wonder announcement. Meta calls them its most capable models yet and the next step on our scaling ladder and a major leap toward personal superintelligence. It also says the series achieves state-of-the-art results across key benchmarks, a claim based on its own testing. Meta is moving beyond lighter Muse Spark models toward larger models for sustained agent work. That means continuing through a task rather than answering one question. Meta trained both at its gigawatt-scale Prometheus data center.
Developers building software and businesses building AI agents are the intended buyers. Wonder is now available through Meta’s API, a paid connection that lets apps send requests to a model, and it’s in preview. Every plan for Muse Code, Meta’s terminal coding agent that works on code from the command line, includes it too. Meta says Wonder will power consumer Muse experiences later, starting with subscribers, but gives no date.
What is Meta Muse Wonder?
Meta Muse Wonder is a pair of AI models that accept text, images, video, audio and PDF files and reply in text. Each can consider up to 1,048,576 tokens at once. A token is a small text unit, sometimes part of a word. The context window limits the material considered in one request.
In its announcement, Meta shows the models building a simple kitchen game, drawing up a wedding seating plan and helping someone work through insurance claims. These are selected demonstrations. The launch materials name no production customers.
Meta describes Active Memory as recalling relevant context across conversations. Active Focus means persisting through blockers, trying alternatives and seeking clarification or confirmation before important actions.
Software already built for OpenAI’s API can talk to Wonder as well. Meta’s developer blog says such tools can point at Meta’s endpoint, the web address that receives requests, and ask for the model ID muse-wonder-1.0. Compatible here means the request format matches, not that every feature works the same way. For local execution, compare Meta’s local Muse Glimmer model. Meta also lists OpenRouter as another route to Wonder.
Wonder Full sits behind a Muse Code subscription
Only two Muse Code plans include Wonder Full: Power, at $50 a month, and Maximum, at $200 a month, according to Meta’s Wonder model page. Wonder itself comes with every Muse Code plan and through pay-as-you-go API access, where customers pay for what they use. Wonder Full is not available through the API, as Meta’s model availability table shows.
Meta’s Muse Code subscription documentation says the plan’s credential is for use with Muse Code only, and any additional API keys are billed at pay-as-you-go rates. Meta advertises 20 times the Everyday allowance for Power Usage and 80 times for Maximum Usage. Usage limits refresh over time, and neither plan offers unlimited use. Meta adds that plan benefits can vary by region and are shown during sign-up.
A company putting Wonder inside its own app needs the API. The model, coding agent, subscription and runtime are separate choices. The runtime is the environment where an agent executes tasks.
What a coding workload costs on the API
Wonder’s API price is $4 per million input tokens, the text sent to the model, and $20 per million output tokens, visible replies and internal reasoning. Input the system has recently stored for reuse, known as cached input, costs $0.20 per million. Grounding, which lets the model draw on web search results, costs $2.50 per 1,000 search queries, and Meta charges no extra premium for very long prompts. Meta’s reasoning documentation adds a less visible cost: reasoning tokens, the model’s internal working steps before it answers, are billed at the output rate.
Claude Opus 5.5 has the same base rates and cache-read price. Creating a reusable cache costs $5 per million tokens for five minutes or $8 for one hour. GPT-6 Sol’s Standard rates are $2 and $10, half Wonder’s. OpenAI identifies GPT-6.1 Sol as a newer model. Teams with lighter workloads can look at Meta’s Muse Spark models, priced at $1.25 for input and $4.25 for output per million tokens.
A hypothetical workday shows how these charges add up. It assumes 100 API calls of 20,000 input tokens each, with 75% served from cache, 2,000 billed output tokens per call including reasoning, and 200 grounding queries. These are illustrative assumptions, not measured performance.
| Part of the bill | Calculation | Cost |
|---|---|---|
| Fresh input | 0.5 million tokens at $4 | $2.00 |
| Cached input | 1.5 million tokens at $0.20 | $0.30 |
| Output, including reasoning | 0.2 million tokens at $20 | $4.00 |
| Grounding | 200 queries at $2.50 per 1,000 | $0.50 |
| Total | $6.80 |
Over 20 identical workdays, that comes to $136. Without any cache hits, the daily bill rises to $12.50. If each call produced another 10,000 reasoning tokens, Wonder would bill 1 million more output tokens, adding $20 and lifting the day to $26.80, almost four times the base case.
Reusing the same code does not ensure a cache hit, so developers should confirm hits through the cached tokens value Meta reports. All these figures are model charges before taxes or outside infrastructure costs. They can’t be set directly against a $50 Muse Code plan either, because the models, usage limits and permitted access differ.
Does Meta train AI on Muse Code submissions?
Meta may use submissions for training when Contributor mode is on, the subscription default. Meta’s Contributor mode help page says Muse Code subscriptions start in Contributor mode, which lets Meta use submitted prompts, code, files, context and responses for training. Users can turn it off through the settings command. Opting out reduces the usage allowance without changing the plan.
Subscribers are told on the same page not to submit sensitive, confidential or personal information to what Meta calls its Discounted Services. Contributor mode is unsuitable for client or internal work containing confidential or personal information.
Opting out only works going forward. Meta’s terms of service say permissions already granted for earlier Contributor-mode submissions remain in place after a user switches off.
API customers get a different arrangement. Wonder’s API offers only Standard mode, and Meta’s terms say it will not use Content submitted in a request to the Standard Services to train Meta Models. One catch applies: if content first sent under Standard mode is later included in a Contributor-mode request, it is treated under Contributor rules.
Not training on data is different from not keeping it. Standard mode does not automatically include zero data retention. Under it, prompts and responses are generally not stored after the reply, with legal and misuse exceptions. That requires separate approval for an organization, support varies by model, and web search is unavailable under it. The ZDR help page does not confirm Wonder-specific eligibility.
What Meta’s benchmark scores leave open
Meta’s Memory Bench scorecard gives Wonder 85.6 and Wonder Full 85.5. This does not establish a real-world memory advantage. In Meta’s internal test, another AI grades recall across conversations of 64,000 to 100,000 tokens containing distractions and updated facts.
Scoring choices lean toward the best case. Meta ran its internal comparisons at the maximum reasoning setting, which spends the most effort on each answer, while outside results keep each evaluator’s own test setup. Some scores are the highest validated or reported result available. Meta’s scorecard cites Artificial Analysis, Data curve and Abundant AI as evaluation sources. Different test setups limit comparisons and do not establish production performance.
UK and EU buyers still need confirmation of sign-up eligibility, local currency billing, taxes and processing locations. Meta’s geographic use policy lists restricted regions, but absence from that list doesn’t guarantee access. One purchasing question remains unresolved. The Contributor-mode documentation does not specify how much usage subscribers lose when they opt out.
