On 12 October 2026, Meta AI Research published Muse Wonder, introducing Muse Wonder and Muse Wonder Full. The post calls them the lab’s most capable models for personal and professional agent tasks, trained at Meta’s gigawatt-scale Prometheus data center, and says the release builds on Muse Spark. The model card lists the same release date, names Meta Platforms, Inc. or Meta Platforms Ireland Limited as the provider, and says the models take text, image, video, and audio in and return text, with a context length of 1 million tokens at inference. The card and the blog do not state a parameter count or an architecture.
Who can call it today
The blog says Muse Wonder is available today through the Meta Model API and Muse Code, and Muse Wonder Full for Muse Code subscribers. Wonder will soon power Muse, subscribers first. Later, Meta plans specialized models it “can bring to everyone using Muse.” The API page repeats that split. The Muse app page did not yet mention Wonder on launch day. The Verge reports the same availability and no timeline for the specialized models.
CNBC adds prices the research post does not list. Alexandr Wang told CNBC the codename was Watermelon. He said: “We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time.” CNBC says Muse app users will need $20 or $100 a month to access the model, and that Wonder Full is only on Muse Code at $50 or $200 a month. Wang said the API costs more than those subscriptions because it is a “much larger model and at frontier level,” and that it will be priced “meaningfully below” leading OpenAI and Anthropic models.
The Muse Code page lists four monthly plans, at $5, $15, $50, and $200. Those lines do not name Wonder or Wonder Full. Its token table lists rates for Muse Spark 1.2 and 1.3 and has no Muse Wonder row. CNBC is the source for limiting Wonder Full to the $50 and $200 plans.
What Meta says the models do
The blog describes two product behaviors. Active Memory carries context across conversations and retrieves details for the task. Active Focus keeps working through blockers, tries other approaches, asks when it needs to, and checks in before important actions. Meta says it evaluates focus with FocusBench and sustained software work with SWE-Marathon. It also points to Harvey’s Legal Agent Benchmark, where a model works from source documents and instructions to produce legal work.
The model card says the results use max reasoning, compared with GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 at max reasoning, and that some competitor scores come from the reporting vendor. On SWE-Marathon it names Abundant AI as the source for every model, says the Astra and Fable entries use Codex and Claude Code, and says that row is not a common-harness comparison. Wonder Full, it says, was reported there at “xhigh” effort. The card gives its full results as a chart, not a table.
The API page says Wonder Full “tops agentic and coding benchmarks, including DeepSWE and SWE-Marathon in independent third-party runs.” That is Meta’s sentence. The card is more specific: Datacurve for some DeepSWE numbers, Abundant AI for SWE-Marathon.
On generalization, the blog says that for a given score on tasks targeted in post-training, Wonder Full does better on tasks it did not target than the Muse Spark 1.3 trend predicts. Meta attributes the shift to larger base models. That comparison is Meta’s.
Where an outside index places them
Wired reports that Meta says the largest version outperforms Claude Fable 5.1 and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index. The models page, opened 12 October 2026, shows Wonder Full (Max) at 56, Fable 5.1 (max with fallback) at 53, Wonder (max) at 52, and, in its top-five sentence, Opus 5.5 (Max, Default Fallback) at 58. Other Opus 5.5 settings on that page are 56 and 54. Wired’s ordering matches Wonder Full against Fable 5.1 and the top Opus setting.
Safeguards, and what is still unpublished
The blog says Muse Wonder “offers our strongest safeguards to date” against chemical, biological, radiological, nuclear, and cyber misuse, under the Superintelligence Scaling Framework. A full report with additional safety evaluations “will be available soon.” The release does not link it yet.
The model card, published with the release, is a separate document. Under the framework, Meta writes that without mitigations it does not rule out the high-risk threshold in the chemical and biological domain, that unmitigated cybersecurity and loss-of-control results were “moderate or lower,” and that with mitigations residual risk is “moderate or lower” across those domains. Meta says that supports deployment. Those sentences are Meta’s assessment. CNBC reports that the safeguards come only with the paid model for now.
CNBC also reports that some users have complained the Muse agent accessed personal messages without authorization, and that Meta disputes the complaints.
The weights are not offered for download.
The Campfire
No commentsNobody has pulled up a log by this one yet. Be the first to say what you make of it.
Held for the desk. It appears after a look.