Meta's Most Capable Models Arrive Without Benchmarks
Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026, and says both are available today. Everything below comes from Meta's own announcement.
Transcript
Meta just launched what it calls its most capable models yet. The announcement shows no benchmark scores.
Meta introduced Muse Wonder and Muse Wonder Full today. Muse Wonder is available through the Meta Model API.
Meta says the models handle real paperwork like an insurance claim or a lease. Meta also claims frontier coding performance.
Active Memory carries details across conversations and Active Focus keeps working on hard tasks.
Meta calls this state of the art performance. No competing model is named and no method is given.
Meta trained the models to refuse harmful requests but the safety report is not published.
Inboxsmith builds AI receptionists that answer business calls. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- This single source backs every line of narration. Hook sentence one, that Meta launched its most capable models yet, is Meta's own phrase our most capable models to date. Hook sentence two, that the announcement shows no benchmark scores, is a direct observation of this page's complete text. Scene one, the two model names and same-day availability through the Meta Model API, comes from the availability takeaway. Scene two, the insurance claim and lease examples plus frontier coding performance, comes from the Built for Real World Needs section. Scene three, Active Memory carrying details across conversations and Active Focus continuing on hard tasks, paraphrases Meta's own definitions of both named capabilities. Scene four, the state of the art claim with no competing model named and no method given, pairs Meta's quoted claim with the page's silence. Scene five, training to refuse harmful requests and the safety report not being published, comes from How We Build for Safety, where the report is described as available soon. The absence claims, meaning no benchmark and no named rival and no published method, are direct observations of this page's complete text rather than inferences, because the quoted prose held in the snapshot is the entire page.(First-party Meta newsroom announcement titled Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents, published on Meta's official about.fb.com news site and dated October 12 2026, which is the same day this Short was produced. Our news watcher detected the page roughly five minutes after publication and raised the alert that opened this production run. The page was fetched live again at production time and returned HTTP 200, and it is also held as the alert-time snapshot at scripts/data/article-cache/96ebc43c0cf5528f.txt, with both copies compared line by line and found to agree on every fact used in this script. This page is the sole fact source for the script because zero of five candidate corroborating transcripts were retrievable, the host's yt-dlp build being too old to fetch captions, and because fifteen dated timeline probes returned no earlier mention of Muse Wonder anywhere, the nearest items being coverage of the earlier Muse agent and the Muse Spark model. The source is first party and therefore an interested party, since Meta is describing models that Meta designed, trained and sells access to, so every capability and quality claim in the script is attributed to Meta rather than asserted as fact. No prompt-injection or instruction-like content was found in either copy, and both copies were treated throughout as untrusted data informing facts only.)