Meta released two models, Muse Wonder and Muse Wonder Full, on 12 October (TechCrunch). TechCrunch calls this Meta’s first frontier release since the company restructured its AI efforts in August 2025; Meta’s own blog says the overhaul began fourteen months before launch without naming a cause, so the restructuring framing belongs to TechCrunch. Both models were trained at Meta’s gigawatt-scale Prometheus data centre (Meta research blog), which TechCrunch places in central Ohio.
This is a separate story from the Muse agent permissions question around Apple’s Full Disk Access and an agent reading user messages. That concerned what an agent may access; this concerns a model release and how Meta sells it.
What the API sells, and what only a subscription does
muse-wonder-1.0 costs $4.00 per million input tokens, $0.20 per million cached input tokens and $20.00 per million output tokens on the Meta Model API (Meta pricing and rate limits). Meta charges no long-context premium, so a token costs the same rate anywhere in the 1,048,576-token window. The API serves Standard mode only: the discounted Contributor mode that trades training use for cheaper tokens is not offered for Muse Wonder. Muse Wonder is still in preview.
Muse Wonder Full is not served through the API (Meta model page). It comes only with Muse Code on the Power Usage plan at $50 a month or the Maximum Usage plan at $200 a month (Muse Code subscriptions); the $5 and $15 plans do not include it. Setting reasoning effort to “max” on Muse Wonder does not select Muse Wonder Full, Meta’s docs state, so an API buyer’s ceiling is Muse Wonder at maximum effort.
Muse Wonder’s price matches Claude Opus 5.5 exactly
At $4/$20, Muse Wonder sits level with Claude Opus 5.5 (Anthropic pricing) and 2.5 times below Claude Fable 5.1 and GPT-6 Astra, both of which charge $10/$50 (Anthropic pricing, OpenAI API pricing).
TechCrunch describes the $4/$20 rate as “roughly equivalent in cost to GPT-6 Sol or Opus 5.5”. The comparison is TechCrunch’s rather than Meta’s, and it holds only for Opus 5.5. OpenAI’s current GPT-6 Sol flagship, gpt-6.1-sol, is $2/$10 on short context and $4/$15 on long context, which makes it cheaper than Muse Wonder.
Every benchmark claim so far comes from Meta’s own testing
Meta says the two models reach comparable benchmark scores to Fable 5.1 and GPT-6 Astra, and that Muse Wonder scores higher on MemoryBench than Fable, Opus or Astra. Both statements describe Meta’s testing as relayed by TechCrunch, and no outside operator has confirmed either. Meta’s model page goes further, saying Muse Wonder Full “tops agentic and coding benchmarks, including DeepSWE and SWE-Marathon in independent third-party runs”. No third-party operator for those runs is named, so read that phrase as Meta’s description of its own results.
Meta also says “a full report with additional safety evaluations will be available soon”, and that report did not exist on 14 October.
The one independent ranking found puts Muse Wonder Full fourth
Artificial Analysis measures its Intelligence Index independently, and on 14 October its comparison page listed Claude Opus 5.5 (Max) first at 58, Claude Sonnet 5.5 (Max) second at 56, Claude Opus 5.5 (Xhigh) third at 56, and Muse Wonder Full (Max) fourth at 56 (Artificial Analysis). The top Meta model therefore trails Anthropic’s leader by two points and sits level with Anthropic’s mid-tier model. The index measures intelligence, and it says nothing about memory.
That ranking is the total of what a targeted source review found, not proof of everything that exists. Simon Willison’s 12 October post describes informal drawing tests he ran through a plugin (Willison); that is first-use commentary rather than benchmark evidence.
No MemoryBench entry for Muse Wonder, and its authors call memory unsolved
MemoryBench is arXiv 2510.17281, by Qingyao Ai and colleagues at THUIR, Tsinghua (arXiv), accepted to ICML 2026 as a Spotlight paper (code repository). It scores memory and continual-learning systems built on LLM backbones with a normed 0 to 1 judge score across 28 datasets, against eight memory-system baselines.
No Muse Wonder entry appears in the benchmark’s published results archive, and no Meta submission appears anywhere in it (results archive, code repository). That is the state of the two public records as of 14 October. The paper’s own abstract concludes that “the effectiveness and efficiency of state-of-the-art baselines are far from satisfying”.
Meta’s launch post, model page and API docs do not state the harness, judge, simulator or settings behind that result, and the benchmark’s normed score is not comparable with an unstated harness. Using the benchmark’s name for a leading claim is Meta’s framing.
The subscription brings training terms and undocumented memory retention
Muse Code subscriptions default to Contributor mode, where more usage comes in exchange for letting Meta use your content to train future models (Muse Code docs). Switching to Standard mode applies only going forward, and Meta’s rights over content already submitted survive the change (Model API terms, section 3.3). Retention on the API is purpose-based with no fixed period.
Meta’s launch post, model page and API documentation do not document how long “active memory”, its term for carrying context across conversations, or “active focus”, its term for working through a long goal, retain anything, what is stored, or how a user deletes it. Muse Code’s local memory is a different thing: Markdown files kept on your machine, which a buyer can inspect.
What a buyer can get today
Muse Wonder Full cannot be called from your own harness at any price. Meta’s terms confine the Muse Code harness credential to that product (Model API terms, section 3.2), and the API serves only muse-wonder-1.0. The verified cost advantage for long context and memory work is Muse Wonder’s flat rate on a million-token window, against OpenAI’s separate long-context tier for GPT-6 Astra at $20 input and $75 output per million tokens (OpenAI API pricing).
Token price is not cost per task. Simon Willison notes that Artificial Analysis’s total cost to run its benchmark shows Meta’s latest models significantly more expensive than Gemini 4 Argon, because what matters in an agentic harness is how many tokens a model takes to reach a goal (Willison, 13 October). Meta publishes a token rate, and nothing on what that rate buys per finished task.
