What is Muse Wonder? Meta's model, Full vs API

What is Muse Wonder? It is the model Meta calls its most capable to date, announced on October 12, 2026 and built for coding and long agent tasks. It comes in two variants. Muse Wonder runs on the Meta Model API as `muse-wonder-1.0`, in preview, and in Muse Code. Muse Wonder Full runs only in Muse Code, on a subscription.

That split is the part launch coverage tends to skip. If you build on the API, you get Muse Wonder, not Full, and no request setting changes that. This post covers what each variant is, where you can call it, what it costs per token, whether it is open source, and how to hand it live data from the web. Every Meta fact below comes from Meta's announcement, Meta's models page and Meta's pricing page, all read 2026-10-13.

What is Muse Wonder, and what is it for?

Muse Wonder is a hosted Meta model for software engineering and long-horizon agent work. Meta's models page says it makes and verifies changes across many files in real repositories, plans, delegates to sub-agents, calls tools in parallel and recovers from errors across many steps. It also drives a desktop through computer use.

The spec, from the same page:

• Input: text, image, video, audio and PDF. Output: text.
• Context window: 1,048,576 tokens.
• Reasoning effort: levels from "minimal" to "max".
• Endpoints: the Responses, Chat Completions and Messages APIs.
• Also handles: image, video and audio understanding, tool calling, structured output and search grounding.

Meta's announcement names two traits. Active Memory is "the ability to carry what matters to you across conversations". Active Focus is "the ability to keep working toward your goal when a task gets hard, try other approaches, recover from errors, and follow through". Meta also says Muse Wonder will soon power experiences in Muse, rolling out first to subscribers.

Meta publishes benchmark results in the Muse Wonder model card, linked from the models page. Those are Meta's own runs. We ran none, so this post quotes no scores.

Muse Wonder vs Muse Wonder Full: which one can you use?

Muse Wonder Full is not on the API. Meta calls Full the higher-capability variant and makes it available only in Muse Code with a subscription. Setting reasoning effort to "max" on the API does not select it, the docs say. On the API, `muse-wonder-1.0` is the only Muse Wonder you can call.

Source: Meta's models page, section "Muse Wonder availability", read 2026-10-13.

Standard mode means Meta does not use your prompts, files and responses to train its models. Contributor mode trades that permission for more usage or a lower price. On the API, Contributor mode is not offered for Muse Wonder. In Muse Code, subscribers can run both variants in either mode, and Contributor mode is on by default.

What does Muse Wonder cost?

On the API, Muse Wonder is priced per token, with no minimum. Meta charges the same rate whether the context window is nearly empty or nearly full.

Web search grounding, the built-in `web_search` tool, adds $2.50 per 1,000 search queries on top of tokens (Meta's pricing page, read 2026-10-13). What one agent task adds up to, with the arithmetic, is in Muse Wonder API pricing.

Is Muse Wonder open source?

No. Meta's open-weight model is Muse Glimmer, distilled from Muse Spark, which you download and serve yourself. Muse Wonder is hosted: you call it over Meta Model API or use it in Muse Code. Meta's docs list no Muse Wonder weights, so a download that claims to be them is not Meta's release.

Muse Wonder or Muse Spark: which should you call?

Meta's own guidance: both take the same requests on the same endpoints, so switching is a model id change. Pick Muse Wonder for long, complex tasks where capability matters more than cost per token, such as multi-file coding changes, multi-hour agent runs and sub-agent orchestration. Pick Muse Spark for high-volume, latency-sensitive or cost-sensitive work, such as sub-agent tasks, classification and tight tool loops. Muse Spark also has a discounted Contributor mode on the API.

How do you call Muse Wonder as a developer?

Meta's examples point the OpenAI SDK at the base URL `https://api.meta.ai/v1` and pass `muse-wonder-1.0` as the model. The key is read from `MODEL_API_KEY`. This lists the models your team can call:

curl -X GET "https://api.meta.ai/v1/models" -H "Authorization: Bearer $MODEL_API_KEY"

Meta's coding agents guide, its agent frameworks guide and its computer use guide all use `muse-wonder-1.0` for long-running agent work.

How do you give Muse Wonder live data?

Muse Wonder can call tools, and it has one built in: search grounding through `web_search`, billed per query. For anything else, a news feed, a page as clean text, a set of search results as rows, you define a tool and your code runs it. The model asks for the call; your app or agent client makes it and hands back the result. How that loop works is in LLM tool calling with live APIs.

looot is one way to fill that tool slot. One key and one prepaid balance give your agent data jobs it can call by name: `job:web.search` to search the web, `job:web.scrape.markdown` to read a page as markdown, and `job:news.search` for today's news. An agent client that speaks MCP can add looot as a server, explained in what an MCP server is. looot works as an MCP server inside Muse Code (Adel tested it on 2026-10-13), and you add it like any other MCP server. The setup is in looot's skill file. Each job's inputs and outputs are on the web search API page and the URL to markdown page.

Paste this into the agent that runs Muse Wonder for you, or into any agent that can read a skill file. It prices each call before it spends and stops after three calls.

The one-line setup is in set up looot with one line.

What it cannot do yet

• Full on the API. Muse Wonder Full is not available on the API, and "max" reasoning does not unlock it.
• A stable API. `muse-wonder-1.0` is in preview, so behaviour and prices can change.
• Every country. Meta says model availability can vary by country and points to its Geographic Use Policy.
• A published safety report. Meta's announcement says a full report with more safety evaluations will be available soon.
• Independent scores. The benchmark numbers are Meta's. We did not run a head-to-head.
• Acting for you. looot reads data and calls provider APIs. It does not post, log in to sites as you or read private accounts, and re-running a brief on a timer needs scheduled runs, not available yet.

Last checked 2026-10-13. Meta facts were read that day from the Meta pages linked above. looot prices in the table are read from the live catalog when the page loads.