Muse Wonder and Wonder Full

The short answer

Muse Wonder is the first frontier class model from Meta Superintelligence Labs, announced on October 12, 2026, alongside a larger sibling called Wonder Full. Wonder is available now in two places: the Meta Model API, at $4 per million input tokens and $20 per million output tokens, and Muse Code, Meta’s terminal coding agent, where launch coverage reports it is included in every plan. Wonder Full has no API at all. It lives only inside the top two Muse Code plans. Neither model is inside the consumer Muse app yet. Meta says Wonder will soon power experiences in Muse, rolling out first to subscribers, and has published no date. This guide separates what Meta claims, what independent testing shows, and what nobody knows yet.

What Wonder is

To place Wonder, start with the model it is not. Muse Spark, the model behind the Muse app since launch, was built lightweight and fast: a personal agent that answers quickly, runs errands across your apps, and keeps its costs low enough to offer on a free tier. Wonder is the opposite trade. It is larger, slower, and far more capable, built for long running agentic work: tasks that stretch across apps and across hours, professional work in fields like law, finance and health, and serious coding. Mark Zuckerberg described it as a much larger and more advanced class of models than the Spark family, and TechCrunch called it what Meta itself calls it: the company’s first frontier model. It is also the first frontier class release from Meta Superintelligence Labs, the research unit led by Alexandr Wang.

Two abilities carry the launch. The first is Active Memory: Wonder carries context from one conversation into the next and retrieves what matters days, weeks and months later. In Meta’s launch video, research scientist Kavya Kopparapu put it this way: “We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond.” The second is Active Focus: Wonder persists through hard tasks instead of giving up at the first failure. It tries alternatives, recovers from its own errors, delegates pieces of work to sub agents, and asks before it takes a consequential action.

Meta also calls Wonder its most aligned model to date. The documented specifications, gathered from Meta’s developer documentation, are these:

• Model ID: muse-wonder-1.0
• Context window: 1,048,576 tokens, just over one million
• Inputs: text, image, video, audio and PDF files. Output: text only
• Reasoning effort: adjustable, up to a maximum setting
• Tools: tool calling, structured output, search grounding and computer use
• Status: in preview in Meta’s developer documentation

The models were trained at Meta’s Prometheus data center in Ohio. None of this makes Wonder a new app. It is a model, and where you can reach it today is a developer story, which is why the access question matters more than the benchmark question for most readers.

Wonder vs Wonder Full

Wonder Full is the bigger sibling, and most of the headline numbers Meta quotes belong to it. If you only read the launch headlines, you have mostly been reading about Full. The difference that decides things in practice, though, is not power. It is access.

Wonder is the model with an API. Any developer with a Meta developer account can call it today and pay per token. Wonder Full cannot be called at all. There is no endpoint, no token price, and no announced plan for one. Full exists only inside Muse Code, and only on the Power Usage and Maximum Usage plans. So the choice between the two models is mostly made by where you work: in your own product, through an API, Wonder is the only option; inside Meta’s coding tool on a top plan, Full is part of the subscription.

The benchmarks, honestly

Meta published a model card alongside the launch, with scores measured at maximum reasoning effort. Read those numbers as Meta’s claims, because that is what they are. On the tests Meta chose to headline, the Wonder models lead:

• DeepSWE 1.1, a coding benchmark: Wonder Full 79.0, Wonder 78.8, GPT-6 Astra 74.1, Claude Opus 5.5 74.2
• SWE-Marathon, long coding tasks: Wonder Full 57.5, Wonder 53.1, GPT-6 Astra 42.5, Claude Opus 5.5 46.3
• Harvey LAB-AA, legal work: Wonder Full 13.1, Wonder 10.0, GPT-6 Astra 8.6, Claude Opus 5.5 4.2
• MemoryBench, memory across conversations: Meta puts both Wonder models at about 85.5

One figure in that list deserves a flag of its own. The MemoryBench score traces back to Meta’s model card as reported by a single secondary source, so hold it more loosely than the rest.

Now the independent check. Artificial Analysis, which runs its own evaluations, puts Wonder Full at 56 on its Intelligence Index and Wonder at 52. Full ties Claude Sonnet 5.5, finishes ahead of GPT-6 Astra and Gemini 4 Argon, both at 53, and behind only Claude Opus 5.5 at 58. Meta’s own summary of its card is more modest than the launch headlines: comparable benchmark scores to Fable 5.1 and GPT-6 Astra.

The pushback arrived within a day. Bindu Reddy, the chief executive of Abacus.AI, publicly challenged Meta’s framing, said Wonder was “nowhere near” its rivals, and promised detailed results of her own. A developer analysis shared widely since argues the sharper point: at identical pricing, Wonder loses 9 of the 14 benchmarks in Meta’s own model card to Claude Opus 5.5. Artificial Analysis adds a cost footnote that matters for the pricing section below: Wonder Full burned about 140 million output tokens running its index, against a median of 88 million for rival models. At maximum effort, Wonder is hungry.

One more caution, because charts travel faster than corrections: a rankings chart that spread widely on launch day was flagged as likely fan made. This guide quotes the model card and the independent index, and nothing else. The fair summary is that Wonder puts Meta in the frontier tier, with a genuine fight on coding and legal work. It is not a clear champion. And both models are closed and hosted only, with no open weights: a real break from the Llama era, when Meta gave its models away, and even from the smaller Muse Glimmer, which shipped as open weights under Apache 2.0.

What Wonder costs

Wonder on the Meta Model API is pay as you go: $4 per million input tokens and $20 per million output tokens. Cached input, the repeated context a long conversation reuses, costs $0.20 per million. The optional web search add on costs $2.50 per 1,000 queries. The rates hold whatever context length you use, short prompt or million token document alike. New to token pricing? Our Muse tokens explained guide walks through how tokens work before the numbers below have to make sense.

That headline price is exactly the list price of Claude Opus 5.5, and the parity is worth pausing on, because list price is not what a task costs. Artificial Analysis measured the average cost of completing one of its index tasks and found Wonder at $4.52, Opus 5.5 at $5.98, and GPT-6 Astra at $3.26. Wonder uses more output tokens to reach its answers, which pulls its real cost up from the sticker. It still undercuts Opus on that measure, but anyone budgeting from the $4 and $20 figures alone will underestimate the bill.

Wonder Full has no token price because there is no API to price. It reaches you as part of a Muse Code subscription: Power Usage at $50 a month carries 20 times the usage allowance of the entry plan, and Maximum Usage at $200 a month carries 80 times. If a report tells you Wonder Full costs $50 or $200 per million tokens, or per anything else, it has confused a plan price with a model price. The sections below keep the two apart.

How to get it today

The Meta Model API

The API lives at api.meta.ai/v1 and speaks the interfaces developers already use: it is compatible with the OpenAI and Anthropic formats, including Responses, Chat Completions and Messages, so most existing tools and code work with a new key and a new base address. Access is pay as you go with an API key under Meta’s model usage terms; commercial use is permitted, and customers own their inputs and outputs. The documented rate limits are 3,000 requests a minute and 4 million tokens a minute per team. Wonder is also listed on OpenRouter as meta/muse-wonder. Our earlier Meta Muse API explainer covered the API as it was announced in September; this guide covers what actually shipped. Remember that Meta’s documentation labels Wonder in preview before building anything load bearing on it.

Muse Code

Muse Code is Meta’s terminal coding agent: you run it inside a project folder on macOS, Linux or Windows, and it works through coding tasks with you. It launched in beta on August 5, 2026, powered by Muse Spark 1.2, and left beta around September 2 with monthly plans. Launch coverage reports that Wonder is included across every Muse Code plan, while Wonder Full sits only in the top two. One practical catch, flagged by people reading the documentation closely: the Muse Code docs still list Spark 1.2 as the default model. If you open Muse Code expecting Wonder, check which model is actually selected, and switch it yourself.

Two plan families, and why people mix them up

Two different products from Meta now carry similar plan names, and launch coverage has repeatedly smeared them into one. Keep them apart and every price in this guide makes sense.

The consumer Muse app is the personal agent on your phone. Its plans are Free, Power at $20 a month, and Maximum at $100 a month. Those plans buy usage inside the app, where the agent still runs on Muse Spark today. Wonder is not in the app yet, at any price. Our Muse for Small Business setup guide walks through the app and its plans as they exist today.

Muse Code is the developer tool. Its plans are Everyday Usage at $5 a month, High Usage at $15, Power Usage at $50, and Maximum Usage at $200. Wonder Full lives in the top two of these. The $50 and $200 figures that early reports attached to Wonder Full are these Muse Code plan prices: not a subscription for the model, and not a token rate.

The tell for bad information is simple. If a price cannot say which of the two families it belongs to, it is probably wrong.

When Wonder reaches the Muse app

Here is Meta’s own wording, from its launch post: “Muse Wonder will also soon power experiences in our personal agent product, Muse, rolling out first to subscribers.” That is the complete official position. Soon, subscribers first, no date.

Update, October 15, 2026: The first subscribers now report Wonder inside the app itself. In a Reddit discussion that began on October 12, 2026, users on paid plans said their Muse answered “Muse Wonder” when asked which model it was running. These are user reports, not a Meta announcement, and the pattern across them is consistent. The rollout looks staged by app version, with reports naming version 7 of the Mac app and version 11 of the iPhone app. Some updated apps still answer Muse Spark 1.3. Every report from a free account still shows Spark. Meta’s official position has not changed: soon, subscribers first, no date. If your app still runs Spark, the reports suggest that is the normal state of a staged rollout, not a problem with your account.

What changes for regular users when it lands, in Meta’s framing, is the model’s two headline abilities arriving inside the agent you already use, the one that works through your connected apps. Active Memory means Muse remembers what matters across days and weeks instead of starting cold. Active Focus means it stays with a hard errand, tries another route when the first one fails, brings in help, and checks with you before anything consequential. Nobody outside Meta has tested the in app version, so treat that as a promise, not a review. If you are preparing for that day, our complete Muse connectors list tracks what the agent can connect to right now.

What is not known about the consumer rollout is a longer list than what is: no date, no country order, no word on whether free tier users receive Wonder at all, and no word on the usage limits that will apply inside the app. The context, reported by The Wall Street Journal, is that download growth for the Muse app has slowed since its September launch, which is one reason the Wonder rollout matters to Meta well beyond benchmarks.

The privacy line developers should know

Developers choosing a route to Wonder are also choosing a data posture, because the two routes treat your inputs differently. The pay as you go API runs in what Meta calls Standard mode, under which Meta does not train on your inputs. For Wonder, pay as you go is Standard only. Approved organizations can go further and get zero data retention, at a price in features: no server side state, no background requests, no file uploads and no built in web search.

Muse Code subscriptions default to the other posture, Contributor mode, under which Meta may train on the code you submit unless you opt out. Meta is not hiding the trade. On the API it even sells a discounted Contributor endpoint for Spark that exchanges training use of your data for a lower price. But defaults are what most people run. If proprietary code is going through a Muse Code subscription, find the opt out before the first real project, not after.

What is not known yet

For the record, the open questions as of October 13, 2026:

• The consumer rollout date, and which countries receive Wonder in the app first
• Any API for Wonder Full. None exists, and none has been announced
• A standalone safety report. Meta has promised one; it had not been published at the time of writing
• Enterprise and volume pricing, and a general availability date for the API, which remains in preview
• Regional specifics for the API preview. Meta has published none

When those answers land, this guide will be updated. Until then, be suspicious of any page that fills them in confidently.

Get started

Frequently asked questions

Is Muse Wonder in the Muse app now?

No. Wonder launched on October 12, 2026 for developers, through the Meta Model API and inside Muse Code. Meta says Wonder will soon power experiences in the Muse app, rolling out first to subscribers, but it has published no date, no country order, and no word on whether free accounts get it.

What does Wonder cost?

On the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 per million and web search at $2.50 per 1,000 queries. Wonder Full has no token price and no API. It is included only in the top two Muse Code plans: Power Usage at $50 a month and Maximum Usage at $200 a month.

Wonder or Wonder Full: what is the difference?

Wonder Full is the larger model, and most of the headline benchmark numbers Meta quotes belong to it. The difference that decides things in practice is access. Wonder has an API and sits inside every Muse Code plan. Wonder Full has no API at all and lives only in the top two Muse Code plans.

Is Wonder better than Claude and GPT?

It sits in their tier, not clearly above it. On the independent Artificial Analysis Intelligence Index, Wonder Full scores 56, level with Claude Sonnet 5.5, ahead of OpenAI’s Astra at 53, and behind Claude Opus 5.5 at 58. Wonder scores 52. Meta’s published model card shows clear coding wins, and critics reply that Wonder loses most of Meta’s own comparisons to Opus at the same list price.

Is Wonder open source?

No. Both Wonder and Wonder Full are closed, hosted only models with no open weights. That is a break from the Llama era, when Meta released its models for anyone to download and run.

How do I try Wonder today?

Three ways. Call it through the Meta Model API with an API key, pay as you go. Use it through OpenRouter, where it is listed at the same published rates. Or open Muse Code on any plan and select Wonder as the model, noting the docs still show Spark 1.2 as the default. Wonder Full needs a Power Usage or Maximum Usage Muse Code plan.

Built with Muse is an independent community project. Not affiliated with or endorsed by Meta.