Muse Wonder and Muse Wonder Full: Access, Pricing, Benchmarks

Two models, two ways in

Meta released two models on Monday, October 12, 2026. Muse Wonder was available that day through the Meta Model API and Muse Code, Meta's coding product launched in August. Muse Wonder Full, the larger variant, was available only to Muse Code subscribers and was not served through the API. Meta said Wonder would also power its Muse personal agent, rolling out first to subscribers, but that part was a plan rather than a launch-day feature. [1][2]

Meta's model card dates both models to October 12, lists a 1 million token context window, and gives text, image, video and audio as inputs with text as output. Meta's API pricing page lists Muse Wonder as muse-wonder-1.0 at $4 per million input tokens, $20 per million output tokens and $0.20 per million cached input tokens, with web search grounding billed separately. TechCrunch put that price roughly level with GPT-6 Sol and Claude Opus 5.5. [3][4][2]

Atlas interpretation: The split matters for anyone comparing Meta with other labs. Developers could build on Wonder, but the variant Meta used for its safety assessment and its strongest claims sat behind a consumer coding subscription, so an API customer could not run the model Meta treated as its most capable. [3][2]

From Spark to Wonder

TechCrunch described Wonder as larger and more capable than Muse Spark, which Meta had designed as a lightweight, low-latency model, and headlined it as Meta's first frontier model, the first such release since Alexandr Wang restructured the company's AI work in August 2025. Meta said the models were trained at its gigawatt-scale Prometheus data center and argued that Wonder Full did better on unfamiliar tasks than the Muse Spark 1.3 scaling trend predicted, which it attributed to larger, stronger base models. The post gives that result as a chart without numbers in the text. [1][2]

The two features Meta named were Active Memory, which carries context about a user across conversations and retrieves what a task needs, and Active Focus, which keeps an agent working through blockers, asks for clarification and checks in before consequential actions. [1]

How Meta scored it, and one outside number

Meta's model card compares Wonder and Wonder Full, at maximum reasoning effort, with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. It mixes sources: some rows come from Artificial Analysis, Vals AI, Datacurve or Abundant AI, some from rivals' own system cards, and some from Meta's internal runs, including an internal composite called the Personal Superintelligence Index and an internal MemoryBench. Meta says its runs of competitor models were best effort and may not be tuned for them, and that its SWE-Marathon row compares models in different harnesses. [3]

TechCrunch reported Meta's claim that the Wonder models scored comparably to Fable 5.1 and GPT-6 Astra and that Wonder beat all three rivals on MemoryBench, and noted these were Meta's own results. Artificial Analysis, testing independently, listed Muse Wonder at maximum effort at 52 on version 4.3.2 of its Intelligence Index. [2][5]

Atlas interpretation: Most of the comparison table is Meta's choice of rows and runs. The Artificial Analysis listing is the only independent measurement cited here, and it covers the API model, not Wonder Full. [3][5]

What the model card says about risk

Meta assessed Wonder Full under its Meta Superintelligence Scaling Framework. Without mitigations, Meta said it could not rule out the model reaching its high-risk threshold in the chemical and biological domain, so it applied safety post-training and system-level classifiers there. It rated the unmitigated model moderate or lower for cybersecurity and loss of control, and said that with safeguards every domain sat at moderate or lower, which supported deployment. [3]

One cyber result shows how much a scoring rule matters. On ExploitGym, Wonder Full captured 92 flags, but a verification pass found only 21 used the intended vulnerability; Meta says 21 is the figure comparable to ExploitGym's published ranking. The card also lists 3D and spatial reasoning as a known limitation, and Meta said a fuller safety report would follow. [3][1]