Guide

Meta Muse Wonder: what's new and how to prompt it

Published 18 October 2026

A home office desk with a laptop, a long checklist pinned to a corkboard, a notebook and a mug of coffee

Short answerMuse Wonder is Meta's new flagship chat model, released on 12 October 2026, built around two ideas: Active Memory, which carries facts across conversations, and Active Focus, which keeps it working on a long task instead of stopping at the first blocker. Give it standing facts once, state the real finish line, and tell it when to check in before acting.

Meta released Muse Wonder, and a more capable sibling called Muse Wonder Full, on 12 October 2026, through its Superintelligence Lab. Muse Wonder is available through Meta's model API; Muse Wonder Full is limited to subscribers of Muse Code, Meta's coding agent product.

Meta built the pair around two named capabilities: Active Memory and Active Focus. Neither is unique to Meta as an idea, but naming them this clearly gives you a useful way to think about how to prompt the model.

What's new

Active Memory. The model is meant to carry facts you've mentioned across separate conversations and bring back the ones relevant to what you're asking now. Meta's own example: you mention in passing that your lease expires in March, and later ask about moving costs; Active Memory is meant to connect the two without you repeating yourself.
Active Focus. This is about sticking with a long or multi-step task: trying a different approach when the first one doesn't work, asking you a clarifying question instead of guessing, and checking in with you before taking an action that matters.
Focus areas. Meta's own description points at coding, and financial and legal questions, as the areas it was tuned for.
Context window. Reported at 1 million tokens, in line with other current frontier models.

Meta's benchmark comparisons (against its own earlier Muse Spark model, and claims of parity with other labs' models) are self-reported, so treat them as a starting point, not a verdict. As with any new model, exact behaviour, pricing and access can shift after launch, so check Meta's own documentation for the current details.

Prompting for Active Memory

The feature only helps if you actually give the model something worth remembering. Two habits make a real difference:

State standing facts plainly, once, not buried in a question. Instead of working a fact into a question where it might get missed, give it its own line:

prompt — 2 blanks

For context, remember this: [STANDING FACT, e.g. my lease ends 15 March 2027, and I have two kids aged 6 and 9].

Now: [YOUR ACTUAL QUESTION]

Ask it to use what it knows, rather than assuming it will. If a new question depends on something you said earlier, point at it directly: "Given what you know about my lease ending in March, what should I budget for moving costs?" This also lets you check whether it actually picked the fact up, instead of hoping.

Because what a chat model remembers across sessions and for how long isn't something you can see directly, it's worth occasionally asking "what do you currently know about me?" to check nothing stale or wrong has stuck around. The same prompts that work for preparing questions for your accountant benefit from this: give your situation once, clearly, rather than retyping it every time you come back to the topic.

Prompting for Active Focus

Active Focus is meant for jobs that take more than one reply to finish; a long project kickoff brief or a one-page business plan are good examples. Three things help it actually behave that way:

Give the whole goal and the finish line up front. Say what "done" looks like, not just the first step:

prompt — 2 blanks

Goal: [THE FULL OUTCOME YOU WANT].
Done looks like: [WHAT A FINISHED, GOOD RESULT INCLUDES].
Work through this step by step. If you hit something you can't resolve, ask me a specific question instead of guessing.

Tell it when to check in. Active Focus is meant to pause before significant actions, but you can make the rule explicit for anything you consider significant: "Before you commit to a final number or send-ready draft, show me your plan first."

Give it permission to try a different approach. If the first method it picks isn't working, say so plainly rather than starting over from scratch: "That approach isn't working because [REASON]. Try a different one instead of repeating it."

This is the same instinct behind scoping a side project you can actually finish or explaining a budget overrun to your boss: give the model the real constraints and the real finish line, and let it work the steps in between, rather than issuing one instruction at a time.

A copy-ready template

prompt — 4 blanks

Context to remember: [STANDING FACTS ABOUT YOU OR YOUR SITUATION].

Task: [WHAT YOU WANT DONE].
Done looks like: [WHAT A GOOD FINISHED RESULT INCLUDES].
Check in with me before: [ANYTHING YOU WANT A GREENLIGHT ON FIRST].
If you hit a blocker: ask a specific question rather than guessing.

Where to go from here

For a similar look at a newly launched chat model, see the Claude Opus 5.5 prompting guide. Browse the chat and writing prompts collection for more ready-made prompts that work across tools, or use the prompt builder to put one together for your own task.

Quick checklist

State standing facts on their own line, not buried inside a question.
Point back at a fact when a later question depends on it.
Give the full goal and what "done" looks like, not just the first step.
Name when you want a check-in before it acts.
Tell it explicitly when to try a different approach instead of repeating one.
Treat Meta's own benchmark claims as a starting point, and check current pricing and access before relying on it.