Meta Muse Wonder: Price, Benchmarks and How to Get It

Meta Muse Wonder is Meta’s first frontier-class AI model family, released on October 12, 2026. It comes in two versions: Muse Wonder, available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, and the larger Muse Wonder Full, which is limited to paying Muse Code subscribers. Independent testing puts Wonder Full at 56 on the Artificial Analysis Intelligence Index, up from 48 for Muse Spark 1.3.

I have been waiting for this one with a mix of curiosity and suspicion. For over a year Meta has been the company that spent like a frontier lab without shipping like one. It hired aggressively, built a gigawatt data center in Ohio, and then released Muse Spark, a quick and pleasant model that nobody mistook for the top of the field. Muse Wonder is the first time the output looks like the budget. Whether it fully earns the word “frontier” is a more interesting question than the launch video suggests, and the answer depends a lot on which numbers you look at.

What Meta actually released

Two models came out of Meta’s Superintelligence Lab on Monday. Muse Wonder is the general-purpose one. It is live in the Meta Model API and inside Muse Code, the coding agent I covered when it arrived in August in our Muse Code explainer. Muse Wonder Full is the heavier sibling, and Meta is keeping it on a short leash: it is not served through the API at all. According to Engadget, you need Muse Code’s $50 per month “power usage” tier or the $200 per month “maximum usage” tier to touch it.

Both models accept text, images, video and audio, and both have a one million token context window. Meta says Wonder will “soon” power the consumer Muse assistant too, starting with paid subscribers, with specialized versions for everyone else later.

Meta’s own description is not shy. Its blog post calls the pair “a major leap toward personal superintelligence,” and Mark Zuckerberg told the Financial Times that Meta has caught up at the frontier.

Muse Wonder at a glance

For reference, Muse Spark 1.3 scored 48 on the Intelligence Index and 54 on the Coding Agents Index, so the jump is real on both.

Active Memory and Active Focus

The two features Meta wants you to remember are Active Memory and Active Focus. They are less about raw intelligence and more about behaving like something you could hand a week-long errand to.

Active Memory keeps the details that matter about you across conversations and pulls the right ones back when a task needs them. In the launch video, research scientist Kavya Kopparapu described the goal as “making sure your model can remember what’s important to you over days, weeks, months, and beyond.” Meta points to a benchmark called MemoryBench, where it says Wonder beats rival models, though that is Meta’s own testing.

Active Focus is about persistence. When a task gets stuck, the model is supposed to try another route, work around the blocker, adapt to new information, and ask before it does something consequential. If Active Memory is a good assistant’s notebook, Active Focus is the habit of not giving up at the first locked door, and also not kicking it down without checking first.

Think about why Meta, of all companies, would lead with these. A model that knows your calendar, your messages and your photos is only useful if it remembers what it learned yesterday. Meta’s demos lean hard on that idea: planning a Saturday run across apps, sorting out an insurance claim, building a wedding seating chart with awkward constraints. It is a pitch for a personal agent, and memory is the part that makes “personal” mean something.

How it scores against Claude, GPT-6 and Gemini

Here is where I slow down. Meta’s model card says Wonder and Wonder Full are “comparable” to Anthropic’s Fable 5.1 and OpenAI’s GPT-6 Astra. TechCrunch noted that Meta gave that comparison without much independent backing, so the third-party results reported by The Decoder are the ones worth reading.

The good news for Meta is solid. Wonder Full lands in the top three on six of ten benchmarks tracked by Artificial Analysis. On the Coding Agents Index it scores 68, tying Claude Code running Sonnet 5.5. On Terminal-Bench it roughly doubled Spark 1.3’s 33 percent while using 30 percent fewer output tokens. On ProofBench, a private math-proof test, it solved 94 of 100 problems, up from 58. And on the Vals AI Index, Wonder Full ranks fifth, ahead of every OpenAI model and behind only current Anthropic ones.

The weaker spots are just as clear. On AA-Omniscience, which measures factual knowledge, Wonder Full scores 30 against 46 for Claude Opus 5.5, 43 for GPT-6 Astra and 42 for Gemini 4 Argon. That is not a small gap. Its hallucination rate fell from 33 percent to 23 percent, which is a welcome drop, but Gemini 4 Argon sits at 15 percent. And for all Meta’s talk of legal and financial expertise, The Decoder reports that finance and legal scores stayed roughly where the predecessor left them, slipping by up to two points.

So my read is this. Wonder Full is a top-tier coding and agent model with a thinner store of general knowledge than its rivals. That is a perfectly respectable place to be. It is not the same as being level across the board.

Pricing: the catch-up is not cheap

Muse Spark 1.3 cost $1.25 per million input tokens and $4.25 per million output. Wonder costs $4 and $20. That is more than triple on input and nearly five times on output.

Per-token prices can mislead, so look at cost per task. Artificial Analysis measured Wonder at $4.52 per task on its index, against $1.60 for Spark 1.3. That makes it cheaper than Claude Fable 5.1 at $7.63 and Opus 5.5 at $5.98, but dearer than GPT-6 Astra at $3.26, Gemini 4 Argon at $1.99 and GPT-6.1 Sol at $0.72. Part of the reason is that Wonder uses 28 percent more input tokens per task, and input makes up about three quarters of the bill.

If you are price-sensitive, this is a strange week to be shopping. Anthropic just went the other direction on its small model, which I wrote about in the Claude Haiku 5.5 price cut piece. Meta is asking premium money while still proving it belongs in the premium aisle.

Safety, and what Meta has not shown yet

Meta says Wonder carries its “strongest safeguards to date” against chemical, biological, radiological, nuclear and cyber misuse, and calls it its most aligned model so far. It also says training focused on resisting adversarial instructions and handling private information carefully. Those are the right things to say for a model designed to read your messages.

The full safety report, however, is listed as “coming soon.” I would have preferred it on day one. A model built around remembering personal details across months deserves its safety evidence at launch, not after the headlines. There is also the older criticism that Muse models are tuned for benchmarks. Meta’s response is that Wonder Full does better on unfamiliar tasks than Spark’s trend line predicted, which it credits to a stronger base model. That claim is testable, and outside labs will test it.

Should you switch to Muse Wonder?

If you already pay for Muse Code, try Wonder Full on a real project this week. The coding numbers are strong enough to justify an afternoon. If you build on APIs, Wonder is worth a benchmark run on your own workload, especially for long agent tasks, but check the bill carefully given how many input tokens it burns. If your work leans on factual recall, such as research or question answering, the Omniscience gap suggests waiting.

The bigger story is simpler. Fourteen months after Meta tore up its AI organization, and with a reported spending plan of up to $145 billion this year, it has a model that sits in the same conversation as Anthropic, OpenAI and Google. A four-horse race is healthier than a three-horse one. You can read Meta’s own account in the official Muse Wonder announcement.

Frequently asked questions

What is the difference between Muse Wonder and Muse Wonder Full?

Muse Wonder is the standard model, available through the Meta Model API and Muse Code. Muse Wonder Full is the larger and more capable version, scoring 56 versus 52 on the Artificial Analysis Intelligence Index. Full is not available through the API and is restricted to Muse Code’s paid tiers.

How much does Muse Wonder cost?

Through the Meta Model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. Meta has not published token pricing for Wonder Full, which you access through a Muse Code subscription at $50 or $200 per month.

Is Muse Wonder better than Claude or GPT-6?

It depends on the task. Wonder Full ties Claude Code with Sonnet 5.5 on the Coding Agents Index and beats every OpenAI model on the Vals AI Index. On factual knowledge it trails Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon by a wide margin.

When will Muse Wonder come to the regular Muse app?

Meta says Wonder will power the Muse assistant “soon,” starting with paying subscribers. It has not given a date. Specialized models built from Wonder are planned for all Muse users later.