Introduction
I have spent the last year watching Meta try to become an AI company again, and I have learned to read its announcements with one eye on the price list. The hard part is knowing whether the thing is actually usable by someone like me — a developer who builds WordPress sites and small tools for clients, not a lab with a research budget.
On October 12, 2026, Meta released Muse Wonder and a larger sibling, Muse Wonder Full, calling them its first frontier models. A month later, on November 9, Meta answered the question I always ask next — where can I actually call it from — in a way nobody expected. Muse Wonder is now served from Google Cloud in preview, and on the same day Meta released Muse Ember, its smallest and cheapest model, built to sit underneath Wonder inside an agent.
That combination is the story: a frontier model you can reach through a cloud you may already use, paired with a cheap worker model that does the boring tool calls for pennies. Let me walk through what Meta actually shipped, what it costs, and where I would be careful before wiring any of it into a real project.
What Muse Wonder Actually Is
Muse Wonder is a large multimodal model from Meta Superintelligence Labs, the division led by chief AI officer Alexandr Wang. Meta calls Wonder and Wonder Full its most capable models, built for personal and professional agentic tasks — from making sense of a lease to building a working app.
Wang was hired in June 2025 and rebuilt Meta's AI effort that August, after the company spent heavily recruiting researchers from Anthropic, OpenAI, Scale AI and other rivals. The earlier Muse Spark line was designed to be lighter and faster. Wonder is the first model Meta is willing to call frontier — a word the industry reserves for systems judged capable of matching the best general-purpose AI on the market rather than trailing a generation behind it. TechCrunch framed it the same way, calling Wonder Meta's first attempt at a true frontier model.
Both models were trained at Meta's Prometheus data center in central Ohio, a site of about a gigawatt that came online in 2026. They were developed under the internal codename "Watermelon," a detail Mark Zuckerberg teased during an Instagram livestream the weekend before launch. Meta rarely confirms codenames before launch, and this one ties the model directly to the Muse consumer app, which had already crossed millions of downloads earlier in 2026.
Meta frames the Wonder series around four capabilities: frontier coding, long-running agents, advanced computer use, and two features it calls Active Memory and Active Focus, meant to let an agent track a task across days or weeks without losing context. That last pair is the one I keep coming back to. Every agent I have built so far has the same failure mode: it starts strong, then forgets what it was doing three steps ago. If Active Memory works the way Meta describes it, that is the difference between a demo and a tool you can leave running. I have not tested it myself, and I would not trust a launch blog on it — but it is the claim worth watching.
The model ID developers call is muse-wonder-1.0. It accepts text, image, audio and video as input and returns text, with a context window of 1,048,576 tokens — enough to hand it a whole WordPress theme, or a client's entire plugin folder, and ask it to reason across the lot instead of feeding it one file at a time.
The Google Cloud Move
Here is the part that surprised me. On November 9, 2026, Muse Wonder became available through Google Cloud in preview, served from Google's partner-model endpoint under publishers/meta on the global location. The independent coverage I read called it the most consequential thing to happen to the model since launch, and I agree.
A developer who already runs on Google Cloud — and that is a lot of us — can now call Meta's frontier model without opening a separate relationship with Meta, without another billing account or API key to manage. It arrives through the same cloud console and the same IAM controls as everything else in the project. That is a distribution breakthrough, and distribution is what decides which models actually get used.
Two things to know before you route production traffic at it: the endpoint is global with no regional alternative, it uses a stateless Responses API with no server-side state, and Meta's terms — not Google's — govern the model.
The price is flat at any context length: $4.00 per million input tokens, $20.00 per million output tokens, and $0.20 per million cached input tokens. Web search grounding adds $2.50 per 1,000 queries. That flat pricing is worth noting, because some providers charge more as context grows. The output side is where the bill lives, as always.
Muse Ember: The Small Model That Matters
On the same day, November 9, Meta Superintelligence Labs released Muse Ember, its smallest and cheapest model. Ember is the first model Meta says was trained using Muse Wonder Full, and it is built to sit underneath Muse Wonder inside an agent — the cheap worker that does the repetitive tool calls while the expensive model does the thinking.
Ember costs $0.10 per million input tokens and $0.50 per million output tokens. That is forty times cheaper than Wonder on input and forty times cheaper on output. Early coverage notes Ember targets cheap agent tool calls and tops the Vals Index on cost.
This is the part of the announcement that actually changes how I would build an agent. The pattern every agent developer converges on is the same: a strong model plans, a cheap model executes. The planning calls are few and expensive; the tool calls are many and should be cheap. Ember is purpose-built for the job, and it is priced to be called thousands of times a day without anyone checking the billing dashboard in a panic.
If you are building an agent that manages a WordPress site — checking for updates, drafting posts, triaging form submissions — Ember is the model that makes the economics work. Wonder plans the migration; Ember runs the hundred small steps. I have not benchmarked Ember myself, so test it on your own tasks before trusting it with anything customer-facing.
The Price Tag, Honestly
Muse Spark 1.3 launched with contributor pricing at $0.10 per million input tokens and $0.20 per million output tokens — a deliberate attempt to undercut rivals by an order of magnitude. Wonder starts at $4 and $20 for the same units. That is roughly a forty-fold jump on input pricing and an even larger one on output. Meta is charging frontier prices to match.
Is it worth it? The honest answer is that the independent numbers are mixed. Artificial Analysis puts standard Muse Wonder at 44 on its Intelligence Index version 5.0.1, ranked 32nd of 74 models, at $6.95 per completed index task and 126 output tokens per second. That is a respectable mid-pack score, not a chart-topper. The first independent comparison arrived on November 2, 2026, when the research firm SemiAnalysis published GDPx, a benchmark of hard agentic tasks across eleven fields. Read it, but still test on your own workload — a benchmark is a proxy for your work, not the same thing.
The other half of the value question is access. Muse Wonder is available through the Meta Model API and inside Muse Code, Meta's AI coding product. Muse Wonder Full is not sold through the API at all — you only get it on Muse Code's $50 or $200 monthly plans, the Power Usage and Maximum Usage tiers. Meta says Wonder will soon power more of the Muse consumer app too, rolling out first to subscribers, with access there tied to a monthly subscription of $20 or $100. So the most capable version is subscription-gated, the standard version is API-priced, and the cheap worker is nearly free — a ladder aimed squarely at developers building agents, not casual chat users.
The Missing Safety Report
There is one gap in the launch I think developers should know about before they build on this model. Meta promised a full safety report for Muse Wonder, but it had not been published as of October 13, 2026, the day after launch. A frontier model without a published safety report is not unusual — they often follow weeks later — but it means the independent safety evaluation that usually accompanies a release of this size was not available at launch.
The other thing to know is that Meta has not announced open weights for either Wonder model. They are served only from Meta's cloud, unlike Muse Spark, whose weights Meta released as Muse Spark Open on October 15, 2026. If your project needs a model you can run yourself, host yourself, or fine-tune yourself, Wonder is not that model. You are renting inference from Meta's infrastructure, on Meta's terms, and if Meta changes the price or the availability, your application changes with it.
What This Means for WordPress and Elementor Builders
The reason I track model releases this closely is that the next generation of WordPress tooling is going to be agent-shaped. WordPress itself is moving that way — the WordPress AI team has proposed native AI agent support for version 7.3, with an MCP adapter, an abilities API and OAuth sign-in — and the models that power those agents will be the ones developers can actually afford to run.
Use Wonder for the planning work: auditing a site, planning a migration, reviewing a theme's code, reasoning across a whole plugin. Its million-token context is genuinely suited to that. Use Ember for the repetitive work: the tool calls, the status checks, the small edits, the form triage. At $0.10 per million input tokens, you can afford to let it check things often. And use the Google Cloud preview if you already run on Google Cloud, because one fewer vendor relationship is one fewer thing to manage — just remember the stateless API and Meta's terms.
Three cautions, from experience. First, test Ember on your own tasks before giving it write access to anything. A cheap model that confidently does the wrong thing a thousand times a day is worse than an expensive model you supervise. Second, keep secrets out of prompts — agent tool calls have a habit of sweeping API keys into places you did not intend. Third, watch the output token bill. Wonder's $20 per million output tokens is where the money goes, and an agent that writes long responses to itself will cost more than you expect.
Conclusion
Meta spent fourteen months rebuilding its AI division around one goal: a model good enough to stop losing developers to Anthropic and OpenAI. Muse Wonder is that bet made public, and the November 9 moves — Google Cloud availability and the Ember release — are Meta making the bet easier to take. A frontier model you can call from a cloud you already use, and a worker model cheap enough to run all day, is a combination I can actually build with.
I am not ready to call Wonder the best model for the job. The independent benchmarks put it mid-pack, the safety report is still pending, and the most capable version sits behind a subscription. But the shape of Meta's lineup is now right: a strong planner, a cheap worker, and distribution through a cloud developers already trust. If you build agents, try Ember this week on a small, supervised task. Try Wonder on a planning job where its context window earns its price. And keep the receipts — test on your own work, because no benchmark, including the ones I have cited here, is a substitute for that.
FAQ
What is Muse Wonder?
Muse Wonder is Meta's first frontier AI model, released on October 12, 2026 by Meta Superintelligence Labs alongside a larger sibling, Muse Wonder Full. It is a multimodal model for coding, long-running agents and computer use, available through the Meta Model API and inside Muse Code. Developers call it by the model ID muse-wonder-1.0.
How much does Muse Wonder cost?
Muse Wonder costs $4.00 per million input tokens, $20.00 per million output tokens, and $0.20 per million cached input tokens, flat at any context length. Web search grounding adds $2.50 per 1,000 queries. The larger Muse Wonder Full is not sold per token; it is available only through Muse Code's $50 or $200 monthly plans.
What is Muse Ember?
Muse Ember is Meta's smallest and cheapest model, released on November 9, 2026. It is the first model trained using Muse Wonder Full, and it is designed to sit underneath Muse Wonder inside an agent, handling repetitive tool calls. It costs $0.10 per million input tokens and $0.50 per million output tokens, and early coverage notes it tops the Vals Index on cost.
Where can I use Muse Wonder?
Through the Meta Model API, inside Muse Code, and — since November 9, 2026 — through Google Cloud in preview, served from Google's partner-model endpoint under publishers/meta. The Google Cloud endpoint is global with no regional alternative, uses a stateless Responses API, and is governed by Meta's terms rather than Google's.
Is Muse Wonder open source?
No. Meta has not announced open weights for Muse Wonder or Muse Wonder Full; both are served only from Meta's cloud. That is different from Muse Spark, whose weights Meta released as Muse Spark Open on October 15, 2026.
Should I use Muse Wonder or Muse Ember for my WordPress agent?
Use both, for different jobs. Wonder is the planner: site audits, migration planning, reasoning across a whole theme or plugin with its million-token context. Ember is the worker: cheap, repetitive tool calls such as status checks, form triage and small edits. Test Ember on your own tasks before giving it write access to anything, and watch Wonder's output token costs.
