Meta launched Muse Wonder and Muse Wonder Full on October 12, 2026, giving developers and small-business software teams a new option for AI coding and multi-step work. Immediate access is through developer tools, including the Muse Code coding assistant.
Everyday Muse users face a different timetable. Meta says Wonder will reach more of its assistant experiences later, with subscribers first; it has not announced a firm rollout date.
Who can use Wonder now?
Muse Wonder Full is limited to two Muse Code subscription plans. Meta’s availability table distinguishes it from Wonder, which also has a pay-as-you-go route.
Model | Access at launch | Published price |
|---|---|---|
Muse Wonder | Meta Model API for use in your own software; Muse Code with pay-as-you-go or any subscription. | API: $4 input, $0.20 cached input and $20 output per million tokens. |
Muse Wonder Full | Muse Code only, on Power Usage or Maximum Usage. No public API access. |
The $50 and $200 plans are Muse Code subscriptions, not consumer Muse plans. They have usage limits and region-dependent benefits and availability. They cover Muse Code usage, not separate API keys used by an app; those are billed separately.
The practical implication: paying for Full in Muse Code does not give an app access to Full. An app using Wonder needs its own API budget.
What the API price means
Tokens are the units used to meter model input and output; cached input is reused prompt content eligible for a lower rate. At the published rates, an illustrative 100,000 uncached input tokens plus 10,000 output tokens costs $0.60: $0.40 for input and $0.20 for output.
If all that input qualifies for cached pricing, the same token totals cost $0.22. This is a calculation, not a measured task: it assumes exactly those billed tokens and excludes search charges, retries and any additional reasoning or output. Real work can require more than one model call.
Check the training setting before submitting work
Muse Code subscriptions start with Contributor mode enabled. Meta’s subscription documentation says this gives more usage in exchange for permission to train future models on submitted prompts, code, files and context, as well as responses.
Turning the setting off excludes future content from training; it does not promise to undo previous contributions. Wonder and Full remain available on eligible plans in either mode. Before sending business material, decide whether that training permission fits the work.
Wonder’s pay-as-you-go API is Standard mode only, which Meta says does not use submitted content for model training.
What Meta means by memory and focus
Meta calls its ability to bring relevant details from earlier conversations into a new task Active Memory. Active Focus is its description of persisting through difficult work, trying another approach and asking for clarification or checking before important actions. These are Meta’s capability claims, not RohitAI test results.
The model card describes MemoryBench as an internal recall test using conversations of 64,000–100,000 tokens, with AI grading. That is narrower evidence than dependable memory over months of everyday use. A launch benchmark cannot establish whether the assistant will reliably remember a particular user’s needs.
For background on the coding product, see RohitAI’s earlier Muse Code and Spark coverage; its older model prices are not Wonder’s prices.
Reporting note: This AI-assisted report draws on Meta’s announcements, current documentation and model card, checked on October 12, 2026. RohitAI did not test the models or complete a subscription purchase.