What the Product Surface Says That the Research Surface Does Not #
Meta shipped a frontier agentic model on October 12, and if you wanted to learn about it from the company’s own research surface, you would find nothing. Muse Wonder is live on OpenRouter, serving real traffic at $4 per million input tokens and $20 per million output tokens, with a million-token context window, multimodal inputs, tool calling, and the ability to operate browsers and desktop applications. The Meta research blog, where the company announced every prior Muse model, has no entry for it. The latest post remains Muse Spark 1.1 from July 9.
The announcement came through a Zuckerberg Threads post and what secondary reporting describes as a Meta research blog entry that is not surfaced on the blog’s main index. Meta did not respond to a request to clarify the announcement path. The product surface is operational. The research surface is empty. The gap is the story.
$4/$20 With No Independent Benchmarks #
The OpenRouter listing confirms the pricing and specs from the vendor directly: $4.00 per million input tokens, $20.00 per million output, $0.20 per million cache read. A 1,048,576-token context window. Multimodal input covering text, images, video, audio, and PDF. Tool calling and structured outputs. Throughput of 71 tokens per second on the Meta provider, 99.99 percent uptime over three days.
OpenRouter’s model card describes Muse Wonder as “a frontier reasoning model from Meta, designed for complex personal and professional agentic tasks” that “plans, delegates to sub-agents, calls tools in parallel, and recovers from errors.” It can also “operate browsers and desktop and mobile applications.” All of these are vendor-reported claims. No independent benchmarks exist for Muse Wonder as of this writing.
Active Memory and Active Focus #
Meta positions two capabilities as Muse Wonder’s differentiators. Active Memory carries context across conversations over days, weeks, and months, and actively retrieves relevant details when they become needed. Active Focus keeps the model working toward a goal through blocks, trying alternative approaches and recovering from errors rather than stopping when the first attempt fails.
Kavya Kopparapu, a Meta research scientist, said the team put “a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond.” Meta claims both capabilities outperform Claude Fable and GPT-6 Astra on internal benchmarks. Those are vendor-reported numbers with no independent verification, and the piece says so plainly rather than treating internal testing as settled fact.
Two Models, Two Access Tiers #
Muse Wonder ships in two variants with different access gates. The API version is available through OpenRouter and the Meta Model API at the $4/$20 per-token pricing. The subscription version, branded Muse Wonder Full, is exclusive to Muse Code and gated behind the Power Usage tier at $50 per month or the Maximum Usage tier at $200 per month. The Everyday Usage ($5) and High Usage ($15) tiers do not include Wonder Full. The subscription version is not available on OpenRouter at all.
Meta also maintains an open-weight sibling: Muse Glimmer 30B, a dense 30-billion-parameter model released August 9 under an MIT-style license, priced at $0.30/$1.20 per million tokens, and optimized for consumer hardware. The frontier model is closed-weight and subscription-gated. The open-weight model is cheap and accessible. Meta runs both tracks simultaneously.
Where This Sits in the Agent Race #
The three-way agent race we mapped in September – OpenAI Dots at $500 per month, Meta Muse as a free consumer app, Apple Siri AI as personal-only – now has a fourth frontier contender. OpenAI charges premium tiers for autonomous agents with 4,000 app integrations. Apple’s Siri AI remains personal-only with no enterprise tier. Muse Wonder splits the difference: a per-token API for developers and a subscription gate for the full model.
The pricing undercuts Anthropic’s flagship and positions above open-weight alternatives. The closed-weight gate means developers building on Wonder depend on Meta’s continued availability and pricing decisions, with no fallback if the terms change. The open-weight Glimmer track softens that dependency for lighter workloads, but the frontier reasoning sits behind a wall.
What Enterprise Deployers Should See #
The practical question for anyone evaluating Muse Wonder is straightforward: can you verify what you are buying? Without a research paper, without independent benchmarks, and without a blog post on the company’s own index, the answer today is that every performance claim traces back to Meta’s internal testing. The product works – OpenRouter is serving real traffic with real throughput numbers. But the gap between the product surface and the research surface is not an accident. It is a choice about where the burden of proof sits.
Meta developed Muse Wonder over 14 months at its Superintelligence Labs under Alexandr Wang, the AI chief hired through a $14 billion Scale AI deal. The lab has the resources to publish benchmarks. The decision not to surface a research entry on launch day, while the product is already serving traffic, tells enterprise deployers something about how Meta wants this model evaluated: by what it does for users, not by what its creators chose to disclose.
Note: All pricing, specifications, and capability claims are from the OpenRouter model listing and secondary reporting as of October 12, 2026. No independent benchmarks for Muse Wonder were found. Meta did not surface a research blog entry on its main blog index at the time of writing.