News · 2026-10-13

Meta launches Muse Wonder, with its strongest version confined to Muse Code

Meta launched Muse Wonder and Muse Wonder Full on October 12, adding a new hosted model family for coding and tasks that require sustained work across tools. Wonder is available through Meta’s developer interface, while the separate Wonder Full is restricted to paid Muse Code plans. Independent testing supports meaningful agent capabilities, with important weaknesses and access limits.

Key facts

• Wonder’s standard rates are $4 per million input tokens and $20 per million output tokens.
• Meta announced the family on October 12, 2026.
• Wonder Full is offered through Muse Code’s $50 and $200 monthly plans, rather than a public per-token interface.
• The primary source is Meta AI Research’s launch announcement, supported by its developer documentation and model card.

The practical question behind the launch is whether another provider can carry a difficult task through changing requirements. A useful coding assistant must do more than answer a programming question: it must inspect files, make changes, test them, remember corrections and recover when its first approach fails. Meta packages that ambition under two names, “Active Memory” and “Active Focus,” in its announcement. Those are descriptions of intended behavior, rather than guarantees about every connected application.

Memory means carrying useful context between conversations and retrieving it when needed. Focus means persisting through blockers, trying alternatives, adapting to new information and checking with the user before important actions. The distinction matters. A huge context window gives a model room to read; it does not automatically make the model select the right old fact, resolve a contradiction or recognize when a task needs a different plan. A filing cabinet can be large while its retrieval system remains poor.

The official model specifications list a roughly one-million-token context window and support text, image, audio and video input, with text output. They also clarify a detail easily lost in launch coverage: raising Wonder’s reasoning effort to its maximum does not turn it into Full. Full is a separate offering. Meta has not disclosed parameter counts or a direct size comparison, so readers cannot infer how much larger it is, or what hardware would run it locally. There are no released weights to download.

Access and cost should therefore be evaluated separately from scores. The pricing page lists cached input at $0.20 per million tokens and says there is no long-context premium. A token-rate comparison does not tell a buyer the price of completing a task: longer reasoning, repeated tool calls and the reuse of cached material all affect the bill. The developer product is live; the consumer announcement describes a future Muse rollout, beginning with subscribers. It does not establish a launch in the Meta AI app.

Meta’s model card reports professional work, browsing, desktop operation, memory and coding evaluations. Its comparisons require care because competitor results can come from different providers or testing arrangements. Meta describes some competitor runs as “best-effort.” The widely circulated 85.6% versus 81.5% comparison concerns Meta’s internal MemoryBench, not a universal intelligence score. It is evidence about one long-conversation recall test, whose grading and setup remain relevant to interpretation.

The strongest outside evidence comes from Artificial Analysis’s October 12 report. It found Wonder Full in the top three on six of ten component evaluations, with strength in agentic knowledge work and scientific reasoning. Factual knowledge remained a relative weakness. Its checked live pages placed Full fourth and Wonder ninth in their displayed rankings. Those changing pages use different displayed cohorts; the ranks are dated observations, not permanent positions or proof that users prefer the models.

Reception mirrors that distinction. The Hacker News discussion welcomed additional competition but challenged whether benchmark scores capture long, iterative coding quality. A brittle answer to an altered familiar riddle circulated on Reddit. That anecdote identifies a possible failure mode, while offering too little evidence for a model-wide ranking. Enthusiasm about provider choice and skepticism about reliability can both be reasonable positions.

The safety disclosure is also conditional. Meta says its mitigations bring residual risks to moderate or lower under its own framework, while not ruling out an unmitigated high-risk threshold in chemical and biological domains. That is a company assessment rather than an external certification. For developers, the useful next step is to test representative tasks, failures and costs under their own agent harness. Meta has supplied another serious candidate; the evidence supports competitive capability without settling the broader frontier claim.

Primary source, verified: read the paper →

Ground Truth. (2026, October 13). Meta launches Muse Wonder, with its strongest version confined to Muse Code. Ground Truth. https://groundtruth.day/news/meta-muse-wonder-launch-access-and-independent-tests.html

@misc{groundtruth:meta-muse-wonder-launch-access-and-independent-tests,
  title  = {Meta launches Muse Wonder, with its strongest version confined to Muse Code},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/meta-muse-wonder-launch-access-and-independent-tests.html}
}

Topics: models · agents · coding · meta · evaluation