News
Meta's Muse Wonder Model: What's New, Pricing, and Who Should Use It
On this page6 sections
Meta Superintelligence Labs released two models on October 12, 2026, and only one of them is available through Meta's API. Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, the same token rates as Claude Opus 5.5, and in Muse Code, Meta's coding agent. Muse Wonder Full, the stronger model in Meta's own tests, runs only inside Muse Code on the $50 and $200 monthly plans. Meta says Muse Wonder will also come to Muse, its consumer agent app, "soon," starting with subscribers.
For a small business owner, Muse Wonder matters mostly through the people and tools that build on it: a developer who picks a model for a booking assistant, or an app that reads documents for you. Two questions decide whether it is worth anything to you: what it costs on your own tasks, and whether Meta trains on what you send. One part of the launch is not yet independently checked: Meta's legal and finance scores for both models. No hands-on test is included; the article works from Meta's announcement and developer documentation and from Artificial Analysis's published scores.
Muse Wonder at a glance
- Released: October 12, 2026, by Meta Superintelligence Labs
- Muse Wonder: model ID muse-wonder-1.0, in preview; Meta Model API, OpenRouter, and every Muse Code plan
- Muse Wonder Full: Muse Code Power Usage ($50 a month) and Maximum Usage ($200 a month) plans only
- API price per million tokens: $4 input, $0.20 cached input, $20 output; no minimum spend and no surcharge for long prompts
- Inputs: text, image, video, audio, and PDF, with a 1,048,576-token context window; output is text
- Independent scores: Full 56 and Wonder 52 on the Artificial Analysis Intelligence Index; Wonder measured at 83 output tokens per second
What is new in Muse Wonder?
1. Two models, and only one in the API
Muse Wonder and Muse Wonder Full come from the same family, but they are reached in different ways. According to Meta's model documentation, Wonder runs on the Meta Model API in Standard mode and on every Muse Code plan, while Full is available only with a Muse Code subscription; setting Wonder's reasoning effort to the maximum does not switch it to Full. Meta's model page also lists OpenRouter as a way to use Wonder in existing tools.
2. Opus 5.5 token rates, and a different cost per task
Wonder's API price matches Claude Opus 5.5 per token and is several times that of Meta's earlier Muse Spark 1.3, which costs $1.25 for input and $4.25 for output per million tokens on Meta's pricing page. Equal token rates do not mean equal bills. In Artificial Analysis's runs of its Intelligence Index, Wonder cost $4.52 per task, against $5.98 for Opus 5.5, which scored 6 points higher, and $0.72 for GPT-6.1 Sol, which scored the same 52. That is a result on one test suite, not a multiplier for your budget: a real bill depends on how much text goes in, how much is cached, and how many steps a task takes. The same test measured Wonder at 83 output tokens per second, slower than Opus 5.5 at 95.6.
3. Memory and follow-through
Meta describes two traits it trained for. Active Memory carries important context about you across conversations and retrieves what the current task needs. Active Focus keeps the model working toward a goal: trying another approach when blocked, asking for clarification, and checking in before important actions. Both are measured on Meta's internal tests, so they show what Meta built for, not how the model behaves on your own work.
What Muse Wonder changes for a small business
Meta's own pitch includes everyday professional work, from "understanding a lease" to making sense of financial information. Its published scores for those areas, all at maximum reasoning effort, look like this:
| Task area (Meta's figures) | Muse Wonder | Muse Wonder Full | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Legal work, strict grading (Harvey LAB-AA) | 10.0 | 13.1 | 4.2 | 8.6 |
| Financial analyst research (Vals Finance Agent V2) | 56.6 | 59.1 | 58.6 | 53.5 |
| Health questions (HealthBench Pro) | 60.5 | 62.3 | 65.6 | 65.3 |
| Business workflows across apps (AutomationBench-AA) | 39.1 | 45.4 | 41.2 | 41.6 |
These are Meta's numbers from its model page; only the Artificial Analysis index above is an independent result.
Reading a lease or a contract
The legal row needs its definition. In Artificial Analysis's version of Harvey's legal benchmark, Harvey LAB-AA v1.1, three AI judges grade each deliverable on 120 private legal tasks, a task earns credit for the share of judges who find every rubric criterion met, and any material hallucination against the source documents sets it to zero. Across the models Artificial Analysis tested, more than 60% of otherwise passing results contained such a hallucination. Wonder's 10.0 is ahead of Opus 5.5 in Meta's table, and it still means most answers fail the strict test. For a salon lease or a supplier contract, use the model to find the relevant clauses and then read each one in the document yourself; GPT-6 Astra, which is covered in a separate article, scores in the same low range.
Numbers and health questions
On financial research, Wonder sits close to Opus 5.5 in Meta's table, and on health questions both Opus 5.5 and GPT-6 Astra score higher than either Muse model. Neither result makes a summary of your monthly figures or a question about a client's health condition safe to use unchecked: compare totals with your books, and take health questions to a professional.
When a developer or vendor builds on it
If someone builds a tool for you, two settings matter more than the benchmark. On the Meta Model API, Wonder runs only in Standard mode, where Meta says your content is not used to train its models. Muse Code subscriptions default to Contributor mode, which gives more usage in exchange for letting Meta train future models on your prompts, code, files, and the responses, according to Meta's subscription terms; it can be turned off in settings at any time. Before client names, invoices, or messages go through a Muse Code session, ask that it be switched off. For business workflows across apps, Wonder sits slightly below Opus 5.5 and GPT-6 Astra in Meta's own table, so a switch is worth testing on your own tasks rather than assuming.
Whichever model reads the document, the decision about what goes back to a client stays with the owner. Empacer is a voice-first AI business assistant for small service business owners. You say what needs doing, review the finished work, and decide whether it goes out.
Who should try it now and who can wait
- You pay a developer to run a tool on Claude Opus 5.5 or GPT-6 Astra: ask for a test of Wonder on your own tasks; the token rates match Opus 5.5, and the API keeps your content out of training.
- Your tool runs many short, repeated tasks: wait or skip; Meta itself recommends Muse Spark for high-volume and cost-sensitive work, and GPT-6.1 Sol matched Wonder's score at a fraction of the cost per task in Artificial Analysis's runs.
- You use AI to read leases and contracts: no reason to switch models for this; with any model, check each clause against the document.
What is published and what is not
| Question | Published by Meta or Artificial Analysis | Not stated in those sources |
|---|---|---|
| Legal scores | 13.1 for Full and 10.0 for Wonder (Meta's model page) | Artificial Analysis's public Harvey LAB-AA leaderboard did not list either model on October 13, 2026 |
| Safety report | "Available soon" (Meta's announcement) | The full report |
| Speed | 83 output tokens per second for Muse Wonder (Artificial Analysis) | A speed figure for Muse Wonder Full, as of October 13, 2026 |
FAQ
Is Muse Wonder free to use?
No free tier is published. The Meta Model API bills per token with no minimum spend, so a few test requests can cost little. In Muse Code, Meta's pages list the Power Usage plan at $50 and Maximum Usage at $200 a month, the two plans that include Muse Wonder Full.
Does Meta train on what I send to Muse Wonder?
Not on the API, which runs Wonder in Standard mode only, according to Meta's pricing documentation. In Muse Code, subscriptions default to Contributor mode, which allows training on your content, and you can turn it off in settings.
Can a developer plug Muse Wonder into existing tools?
Usually. Meta's quickstart points existing OpenAI SDK compatible clients at the Meta Model API, and OpenRouter also offers the model. Limits for muse-wonder-1.0 are 3,000 requests and 4 million tokens per minute, shared across all keys in a team.
How do Muse Wonder and Muse Spark differ?
Meta recommends Wonder for long, complex tasks where capability matters more than cost per token, and Muse Spark for high-volume, latency-sensitive, or cost-sensitive work. Both take the same requests on the same endpoints, so switching between them is a change of model name.