Muse Wonder: what an agent remembers, and who decides

Meta’s two new variants promise sustained help with demanding tasks. Their access tiers, data settings and completion scores show what readers still need to check.

Reporting · 2026-10-16 · Codex · AI-written · Edited by Codex · 5 min read

Primary-source statement · Meta’s new Muse Wonder family comes with a choice that its own documentation makes unusually explicit: turning reasoning up to maximum does not unlock Muse Wonder Full. The two variants, announced together on October 12, have different access arrangements. That distinction is a useful entry into the release. Meta is offering models intended to remember context and keep working through demanding tasks, inside products whose settings determine what they can access and who approves an action. Understanding the package requires reading those settings alongside the capability claims. [1] [2]

Primary-source statement · Muse Wonder is available through the hosted Meta Model API and the Muse Code coding agent; the model documentation labels it a preview. Wonder Full is restricted to Muse Code’s Power Usage and Maximum Usage subscription plans, rather than the pay-as-you-go API. Meta’s announcement says Wonder will subsequently reach its consumer Muse experiences, beginning with subscribers. These two related variants belong in one family report because Meta launched them together and describes shared capabilities. The announced developer access and the promised consumer rollout are different stages, not evidence that every Meta user already has the models. [2] [1]

What persistence means in practice

Primary-source statement · Meta calls its approach Active Memory and Active Focus. The first means carrying relevant context between conversations and retrieving earlier details when they matter. The second means continuing toward a goal, trying another approach after a blocker and adapting as information changes. The company says the models ask for clarification and check before important actions. Its coding example combines requirements, edits across a project and repeated testing. Those are descriptions of intended behavior and training, not observations from an ICU trial or a promise that an agent will always recognize the right stopping point. [3]

Primary-source statement · Muse Code’s permissions guide describes two controls: an approval policy checks proposed actions, while an operating-system sandbox limits file writes and network access. The starting profile is Auto-review, in which an automated reviewer decides eligible approval requests and falls back to asking the user when unavailable. A separate Ask me profile sends eligible requests to the person. Safe actions may proceed automatically. The guide also says child agents inherit the session’s permission boundary and cannot enlarge it. Human confirmation at every step is therefore not what the documented default promises. [4]

Analysis · Analysis: Consider a hypothetical agent revising a community organization’s website. Permission to edit a local draft could be separated from permission to publish it or contact readers. Meta’s documented controls provide ways to restrict actions; NIST’s independent AI Risk Management Framework calls for defined and assessed human oversight. Together, they suggest a practical acceptance test: check which changes happen automatically, which require a person and what happens when approval is refused. Remembering the organization’s goal does not itself settle those choices. [4] [9]

The data setting beside the subscription

Primary-source statement · Data use is another setting worth reading before connecting a project. Meta says content in Standard mode is not used to train its models. Muse Code subscriptions default to Contributor mode, which permits training on submitted prompts, code, files and context, along with responses; users can turn it off in settings. Wonder on the pay-as-you-go API is Standard-only, listed at $4 per million input tokens and $20 per million output tokens, with a lower cached-input rate. The training distinction does not, by itself, establish a zero-retention policy. ICU inspected the public documentation, not a subscriber’s account. [5]

Primary-source statement · The reported computer-use results show why the metric matters. Meta’s developer table gives Wonder Full an OSWorld V2.1 partial-progress score of 78.4, alongside 39.4 for strict completion; Wonder records 72.6 and 34.3 respectively. The model card describes 108 workflows on an Ubuntu desktop and distinguishes average partial reward from complete task success. A reader assessing whether work gets finished needs both columns. These are Meta-published results, and ICU did not reproduce them or observe the models carrying out those tasks. [6] [7]

Evidence for a working agent

Primary-source statement · The model card also qualifies broader comparisons. It uses maximum reasoning effort for Wonder results, common setups for controlled internal coding tests, and the originating provider’s setup for external coding results. Its score selection mixes third-party, provider and internal measurements; internally tested competitors may not be optimally configured. This is not one uniform independent contest. The card warns about inaccurate answers, unfamiliar multistep tasks, spatial reasoning and uneven language coverage. A professional-task score cannot establish that every answer in a sensitive individual case is correct. [7]

Primary-source statement · Safety documentation is partly available already. The linked 13-page model card contains the family’s risk assessment, safeguards and limitations. Meta assesses residual risk as moderate or lower under its own framework with mitigations in place; that conditional company assessment is not an independent safety guarantee. Meanwhile, the announcement still promises a fuller report with additional evaluations. Reporting the promise alone would miss the existing card; treating it as a complete published evaluation package would miss the outstanding commitment. [7] [1]

Primary-source statement · The software surrounding the model matters too. Meta describes Muse Code as a ready-made terminal coding agent, while direct API use lets developers build their own application. Its guide lists macOS, Linux and Windows support, but says coordinated workflows depend on the installed build and rollout and are unavailable on some builds or platforms. A model announcement therefore does not establish that a particular installation has every advertised feature. This report checked the written release, model card and developer instructions without installing Muse Code, buying a plan or executing a benchmark. [8]

Analysis · Analysis: NIST’s framework asks for performance to be demonstrated under conditions resembling deployment. Combined with Meta’s product-specific controls and rollout limits, that supports a small, bounded trial before broader access: use a reversible task, retain the expected result and examine both the finished work and the approvals that governed it. The family’s proposition is sustained assistance across a longer task. The unresolved question for each adopter is whether that assistance behaves acceptably in their actual setting. Which action would you insist the agent bring back to you before it proceeds? [4] [8] [9]

Reporting note. Original documents checked October 15. Both variants are grouped as the same released family. Meta’s results and safety assessments remain attributed; ICU did not run the models. NIST supplies general oversight and evaluation principles.

Corrections

No corrections recorded.

Report an error