Meta releases Muse Wonder and Muse Wonder Full, its first frontier-class models

Meta published the launch at about 16:55 UTC on Oct 12, 2026 through its research blog, newsroom, developer site and posts by Mark Zuckerberg and Alexandr Wang. Zuckerberg wrote: “It’s been just over a year since we started Meta Superintelligence Labs. Muse Spark was our first family of models, and now Muse Wonder is a much larger and more advanced class of models that brings us to the frontier.” The blog calls the pair “the next step on our scaling ladder and a major leap toward personal superintelligence”.

Muse Wonder is the model developers can call: muse-wonder-1.0 on the Meta Model API, in Muse Code, and on OpenRouter. Muse Wonder Full is a separate, higher-capability model, not a reasoning setting (“setting reasoning effort to ‘max’ on Muse Wonder doesn’t select it”, the docs say), and is served only inside Muse Code on the two top subscription plans. The consumer Muse agent does not run on Wonder yet; Meta says it will “soon”, for paying subscribers first.

Meta frames the models around personal and professional agents. Besides the benchmark table (see key facts), the blog argues that Wonder Full generalizes better than its scores imply: plotted against Muse Spark 1.3 at several reasoning efforts, it does better on tasks Meta did not target in post-training than the Spark trend predicts, which Meta attributes to “larger, stronger base models”. The text names “benchmaxxing” and “big model smell” as the complaints this is meant to answer.

Meta’s table shows a mixed picture rather than a sweep. Wonder Full leads on Meta’s internal composite, legal (Harvey LAB-AA), finance, business workflows, memory and long-horizon coding (DeepSWE 1.1, SWE-Marathon). Claude Opus 5.5 leads on health, JobBench, computer use (OSWorld), Terminal-Bench, ProgramBench (70.5 vs 59.0) and the Artificial Analysis index; GPT-6 Astra leads GDP.pdf.

Independent first results agree on the broad placement. Artificial Analysis gives Wonder Full 56 and Wonder 52 on its Intelligence Index v4.3, and notes that knowledge recall (AA-Omniscience) remains weak while hallucination rates are low. Vals AI ranks Wonder Full fifth on its index, behind only Anthropic models.

On safety, the model card reports that the unmitigated Wonder Full might reach the “high risk” threshold of the Meta Superintelligence Scaling Framework for chemical and biological misuse, so Meta applied safety post-training and system-level classifiers; it rates cybersecurity and loss of control “moderate or lower”. The card lists refusal and attack-success figures but not the methods in detail, and Meta says a fuller safety report will follow. Casey Newton remarked on that sentence: “just what you want to hear from Meta on a day that it claims to have reached the frontier of AI”. WIRED noted that Meta “did not disclose whether it has seen agents powered by Muse Wonder misbehave in testing”. Wang told CNBC the team monitors agent “incidents or issues” in Muse, and said of complaints that the agent read messages without permission that users “may have forgotten” earlier authorizations.

The Financial Times ran an interview with Zuckerberg the same day. It is paywalled and was not read for this entry; according to Techmeme’s summary he says Meta has caught up with rivals at the frontier and rejects calls for a slowdown. Zuckerberg and Wang also recorded a two-hour episode of John Collison’s “Cheeky Pint”.

Meta had been absent from the top tier since the Llama line fell behind; Muse Spark (April 2026) was positioned as a lightweight model. Fourteen months after the reorganization under Wang, Meta now has a model that outside evaluators place among the top few, alongside Anthropic, OpenAI and Google. It also keeps its strongest model out of the API, tying it to a subscription product.

The release lands while Anthropic and OpenAI talk about pacing frontier development. Andrew Curran’s widely read reaction was that “Pacing is out the window”, since one large US lab that does not slow down pushes the others to accelerate. Like Muse Spark, neither model has open weights.

• Meta’s benchmark table mixes third-party, provider-reported and internal runs; only the Artificial Analysis and Vals AI results are independent so far.
• The FT, Bloomberg, Axios, NYT and WSJ articles were not readable (paywall or blocked); their content here comes from headlines, Techmeme summaries and quotes in other outlets.
• No knowledge cutoff, parameter count or maximum output length is published.