What happened?
This page is our recap, not Meta's post.
On 12 October 2026 Meta shipped two models, Muse Wonder and Muse Wonder Full. Mark Zuckerberg called them a new family for personal and professional agent work, a bit over a year after Meta Superintelligence Labs started. The newsroom dates the rebuild to fourteen months ago.
You can call Muse Wonder today. Docs call it a preview. On the Meta Model API the id is muse-wonder-1.0. Context is 1 million tokens. It takes text, image, video, or audio, and it answers in text. Price is $4.00 per million input tokens, $0.20 if the input is cached, and $20.00 output. An existing OpenAI-style client can point at that API. OpenRouter is listed as another way in.
Muse Wonder Full is the larger one, id muse-wonder-full-1.0. It is not on the token menu. It comes with Muse Code, on the Power plan at $50 a month and the Maximum plan at $200. Turning Wonder's reasoning effort up to max does not switch you onto Full. The consumer app, Muse, gets Wonder later, and subscribers go first.
Weights are not in this release.
Why this is interesting
It is meant to keep a job going after the chat ends. Meta's names for that are Active Memory and Active Focus. Memory means carrying what matters across conversations and bringing it back when the task needs it. Focus means staying on a hard goal, trying another path, and checking in before anything important. On their own MemoryBench, long chats of 64k to 100k tokens, Wonder scores 85.6 and Wonder Full 85.5. Claude Opus 5.5 is at 81.5 and GPT-6 Astra at 81.4. That bench is Meta's. The gap is still the point of the product.
The long coding jobs are where the card is strongest. On DeepSWE 1.1, Datacurve's third-party runs put Wonder Full at 79.0 and Wonder at 78.8, against Opus 5.5 at 74.2. Astra's 74.1 and Claude Fable 5.1's 67.4 on that row are the labs' own reported numbers, not the same run. On SWE-Marathon, Abundant AI has Wonder Full at 57.5 and Wonder at 53.1, against Opus at 46.3. Meta notes Wonder Full was scored at extra-high effort, and the Astra and Fable figures on that row used Codex and Claude Code. Same bench, not the same harness.
The general index is a different picture. Artificial Analysis measured Wonder Full at 56, fourth on its leaderboard, not an estimate. Their leaderboard lists Claude Opus 5.5 (max) at 58, then Sonnet 5.5 (max) and Opus 5.5 (extra-high) tied at 56 with Wonder Full. Wonder is at 52. Front pack. Not the leader.
Professional work splits by the job. On Harvey's legal-agent bench, the Artificial Analysis version, Wonder Full is at 13.1%, Astra at 8.6%, Opus at 4.2%. The metric is the share of tasks that pass every check, including hallucination checks. The lead is real. The bar is still low. Finance is a photo finish: Wonder Full 59.1, Fable 58.9, Opus 58.6. Health goes the other way: Opus 65.6, Astra 65.3, Wonder Full 62.3. Meta's Personal Superintelligence Index, their own mix of seven tests, has Wonder Full at 59.1 and Opus at 55.2. Read that as their summary, not as an outside ranking.
Closed, priced, and already something you can wire in. This is not a weight drop. The fast model is an API. The full model is a coding subscription. For a company putting agents into software it already runs, the useful facts are the $4 / $20 sticker, the 1 million token window, and a person still confirming anything you cannot undo. Meta says that with mitigations on, residual risk on cyber, chemical and biological, and loss of control sits at moderate or lower. A fuller safety report is still coming.