Muse Wonder arrives for builders before the wider Muse rollout

An assistant remembers last week’s brand brief, but this week the client wants something different. Remembering more only helps if the assistant can recognize which instruction now matters. That is a useful question to bring to Meta’s October 12, 2026 Muse Wonder release.

Meta announced Wonder through the Meta Model API and Muse Code, while Wonder Full is available to Muse Code subscribers. The broader Muse rollout is still described as forthcoming, first for subscribers. Original access announcement.

Memory needs to follow a change

Meta calls the retention and retrieval of relevant cross-conversation context Active Memory. Its Active Focus description covers working through blockers, adapting and asking before important actions. These are Meta’s descriptions of the models; they do not establish how reliably either behavior will work in your project. Research explanation.

Here is an illustrative evaluation for a creator: first supply a short brand brief, then change one important requirement before asking for a draft. Keep the new requirement explicit. Review whether the draft follows it and whether the assistant explains any conflict with the old brief. A fluent response that carries the wrong preference forward would not satisfy that check. We have not run this experiment.

Give the task an ending too. If you want a draft for review, say what should happen when a required source or asset is missing. Continuing through a difficult task is useful only within the job you gave it.

Check the surface, not just the family name

Wonder, Wonder Full and the Muse consumer experience have different access statements. Use the developer entry point and your current account’s model options to establish what you can use. The announcement does not give a universal consumer rollout date; we have not verified a country-specific entitlement or subscription quote.

Before a long task, check the account terms and run one bounded comparison with your existing workflow. For the separate question of how to end a blocked agent task, see our Anthropic evaluation coverage.