How do you use Muse Wonder?

Muse Wonder is the model Meta released a week ago for agents, AI systems that work toward a goal over many steps using tools such as a browser, files or a code terminal. It is not yet in Meta's consumer apps: Meta says it will power its Muse personal agent soon, starting with subscribers. For now, the two doors are built for developers and technically curious users.

The good news is that both doors are straightforward. The trick is in how you hand Muse Wonder a task, because Meta has tuned it to keep going on long jobs and to pause before important actions. Prompts that spell out the goal, the boundaries and the checkpoints get the most out of that design.

The two ways in

A token is a small chunk of text, roughly part of a word. Meta's pricing page adds two details worth knowing. There is no surcharge for very long prompts, so the price per token is the same whether you send a paragraph or a large document. And the cheaper "Contributor mode," where Meta may train on what you send in exchange for lower prices, is not offered for Muse Wonder, so you pay the standard rate.

If you add the web search tool so the model can look things up, Meta charges $2.50 per 1,000 searches on top of tokens.

Worked tasks

Each task below shows how to frame a request so Muse Wonder can plan, persist and check in at the right moments.

Task 1: a long coding job in Muse Code. Meta built Muse Wonder to sustain work over long software projects, and it measures this with a test of long-running coding work. A good request names the goal, the definition of done and the off-limits areas.

Goal: move this project's date handling from the "moment" library
to "date-fns".
Done means: all tests pass, no imports of "moment" remain, and the
bundle size is smaller than before.
Do not change the public API of src/dates.ts.
Work in small commits. Stop and ask me before deleting any file.

The last line puts a checkpoint exactly where an error would be costly.

Task 2: a document question through the API. Meta highlights expertise in law, finance and health, such as making sense of a lease. Ask for the answer and the evidence.

Here is a residential lease (attached). List every clause that
mentions fees or penalties. For each one, quote the clause number,
explain it in one plain sentence, and flag anything that seems
unusual. If a clause is unclear, say so instead of guessing.

Here is a rough cost, with illustrative numbers: a 30,000-token lease plus a 2,000-token answer comes to about $0.12 for input and $0.04 for output, or roughly 16 cents per lease at list prices. Reasoning tokens, the model's hidden working, also count as output, so real bills can run higher.

Task 3: an errand with a checkpoint. Meta's demonstrations include planning a running route and building a wedding seating chart from tricky rules. The same pattern works for any task with constraints.

Build a seating plan for 40 guests at 5 tables of 8.
Rules: couples sit together; Sam and Alex must not share a table;
the two grandparents sit near the exit; children sit with a parent.
If the rules conflict, show me the conflict and two options.
Do not message anyone; give me the final plan as a table.

Asking it to surface conflicts, rather than quietly bend a rule, plays to what Meta calls Active Focus: working through obstacles while asking for clarification when it matters.

A word on early days

Latent Space reported a launch-day bug that used up Muse Code subscribers' allowances too fast; Meta fixed it and reset quotas the same day. Meta also says a fuller safety report is still to come. Keep approvals on for anything that sends, buys or deletes until you have watched the model work.

How does Muse Wonder Full differ?

Muse Wonder Full is the full model in the family; Latent Space's launch roundup describes plain Muse Wonder as the fast frontier version. The practical differences are access, speed and results.

Access. Wonder Full is only available to Muse Code subscribers. It has no published API price, so you cannot call it from your own app or calculate a per-token cost.

Results. Independent testers place Full ahead. On Artificial Analysis's Intelligence Index, a combined score across ten tests, Wonder Full scored 56 and Wonder 52, as reported by Latent Space. On that firm's coding-agent ranking, Wonder Full inside Muse Code scored 68 against 63 for Wonder. Vals AI ranked Wonder Full fifth on its index and Wonder tenth.

Speed and effort. Wonder is quicker: Vals AI measured about 37 minutes per task for Wonder against 57 for Full. Full read fewer input tokens per task in Artificial Analysis's runs, while both wrote about the same amount of output.

A reasonable rule: use Muse Wonder through the API when you are building something or need predictable per-token costs, and reach for Wonder Full in Muse Code when a hard coding task has stalled and the extra time is worth it.