Meta Puts Muse Wonder Full Behind a Code Subscription

Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, 2026, calling them its most capable models for personal and professional agent work. Muse Wonder is live on the Meta Model API and in Muse Code, while Muse Wonder Full is not on the API at all.

The people who can run the stronger model today are Muse Code subscribers on the $50 Power Usage plan or the $200 Maximum Usage plan. Meta says it will later distill specialized models for everyone on Muse. That later step has no date.

The Stronger Model Stays Inside Muse Code

Meta’s developer docs split the family in plain language. Muse Wonder, served as muse-wonder-1.0, is on Model API in Standard mode and on every Muse Code plan. Muse Wonder Full has no pay-as-you-go endpoint. Setting reasoning effort to max on muse-wonder-1.0 does not switch you onto Full.

Chief AI Officer Alexandr Wang called Wonder “our largest and most capable model yet.” Mark Zuckerberg, posting the launch, told people they could put Wonder in Muse by subscribing, and that Meta would try to bring that level of intelligence to everyone over time.

Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work. pic.twitter.com/NKHrybrGKQ

— Mark Zuckerberg (@finkd) October 12, 2026

TWO MODELS, TWO DOORS

Neither model ships downloadable weights. Muse Glimmer remains the Apache 2.0 model you can run on your own machines. Wonder is in preview. Both take text, image, video, and audio, and both write text back, with a 1 million token context window.

The access split is the product. Teams that want Full inside their own stack have no documented way to call it. They can only sit inside Meta’s coding agent.

What Muse Wonder Costs on the API

On Model API, Standard-mode Wonder is billed at $4.00 per million input tokens, $0.20 for cached input, and $20.00 per million output tokens. That card matches the public Opus 5.5 rates and undercuts several $10/$50 rivals. Contributor mode, which discounts Spark 1.3 to $1.25 / $0.15 / $4.25 in exchange for training rights, is not offered for Wonder on the API.

Muse Code is the other till. Everyday Usage is $5 a month and High Usage is $15. Full starts at Power Usage. Subscriptions default to Contributor mode, which grants more usage if Meta may train on prompts, code, files, context, and model replies. Users can switch that off in settings. Future content then stays out of training unless they turn it back on.

MUSE CODE MONTHLY PLANS

• Everyday Usage: $5 a month for Muse Spark, Muse Wonder, and Muse Image, with 10-50 prompts every 5 hours.
• High Usage: $15 a month, at 5x Everyday usage, still without Muse Wonder Full.
• Power Usage: $50 a month, at 20x Everyday usage, and the first tier that includes Full.
• Maximum Usage: $200 a month, at 80x Everyday usage, with the same Full access.

Meta’s Power Usage and Maximum Usage plans are the only listed path to Full. A subscription is a flat monthly rate through the Muse Code CLI, not a token meter you can point at your own product. Pay-as-you-go still exists for Wonder if you bring an API key. Full does not ride along.

Long Coding Runs Are Meta’s Real Lead

Artificial Analysis scored Muse Wonder Full at 56 and Muse Wonder at 52 on Intelligence Index v4.3 at max reasoning effort. Muse Spark 1.3 sat at 48. The gap from Spark 1.3 to Full is 8 points, and the two Wonder releases landed about 6 weeks apart. Full improved on all ten index tests. The largest move was Terminal-Bench 4.0, which went from 33% to 57%. Wonder itself reached 52% on that test.

On the Coding Agents Index, Full hit 68 at max effort inside Muse Code, level with Claude Code running Sonnet 5.5 and ahead of Opus 5.5 at 66. Wonder scored 63. Spark 1.3 on the same index was 54. Abundant, which helped run Meta’s long-horizon coding evals, put Full at 57.5% on SWE-Marathon and first on that board. Meta’s own table lists DeepSWE v1.1 at 79.0 for Full and 78.8 for Wonder, against 74.2 for Opus 5.5.

INDEPENDENT SCORES AT MAX EFFORT

The cheaper API model is also the slower meter. Artificial Analysis put Wonder at $4.52 per index task, 2.8x Spark 1.3’s $1.60, even though output fell to about 54k tokens from 60k, a drop of about 10%. Input use is the bill: 7.69 million tokens per task. Full has no token price, so the lab published no cost-per-task figure for it.

Knowledge is the soft spot. On AA-Omniscience, Full scored 30 and Wonder scored 25, flat with Spark 1.3 and well behind several rivals. Full’s 5-point gain on that test came from cutting its hallucination rate from 33% to 23%. Wonder sits at 27%. Developers arguing over “SOTA” on the launch chart are mostly arguing past this split: long agent coding moved; factual recall did not.

Memory Across Chats and Focus Across Hours

Meta’s research post hangs the product on two named behaviors, Active Memory and Active Focus. Active Memory is meant to carry what matters about a user across conversations and pull it back when a new task needs it. Active Focus is the claim that the model keeps working through blockers, tries other routes, asks when it is stuck, and checks in before a consequential action. FocusBench and SWE-Marathon are the tests Meta cites for that stamina.

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.

Kavya Kopparapu, research scientist, in the launch video

Launch demos stay inside that brief. One clip has Wonder build a playable 3D kitchen game from custom images. Another has it plan a Saturday run in the browser, using personal constraints and OpenStreetMap data. Other spots show it designing swag from a real Instagram profile and building a wedding seating chart from a knot of matching rules. Those are Meta’s chosen tapes, not a random day’s work.

Coding is where Meta says the two behaviors meet: track a spec, edit many files, test, revise, and keep going. Docs describe sub-agents, parallel tool calls, and recovery from errors. That is also why Full living only in Muse Code matters. The long-run scores were posted from inside that harness.

Prometheus Trains the First Frontier Models

Meta said the Wonder models were trained at Prometheus, its gigawatt-scale cluster. In a February 9, 2026 engineering note, the company said Prometheus will deliver 1-gigawatt of capacity once complete, spanning several buildings and tens of thousands of GPUs. Epoch AI places the operational site in New Albany, Ohio.

Zuckerberg said it had been just over a year since Meta Superintelligence Labs started. The research post uses a longer clock: 14 months of rebuilding the stack from first principles, then scaling pretraining and reinforcement learning. Spark was the first closed Muse family. Wonder is the first Meta is willing to call frontier.

THE PATH TO THIS RELEASE

1. Fourteen months before October 12, 2026: Meta says it began a ground-up overhaul of its AI work.
2. February 9, 2026: Meta publishes the Prometheus networking design and the 1 gigawatt capacity target.
3. September 8, 2026: The Muse personal-agent app opens in the United States.
4. September 18, 2026: Muse reaches Canada, 10 days after the US launch.
5. October 12, 2026: Muse Wonder and Muse Wonder Full ship on the API and in Muse Code.

The company that used to lead with open Llama weights is now selling its peak model only through its own cloud and a terminal agent. Glimmer still covers the downloadable lane. Wonder does not.

US and Canada Wait for Wonder in Muse

The consumer app is the other queue. Meta says Wonder will soon power experiences in Muse and will roll out first to subscribers. Free users get no date. The app itself is still a two-country product: the United States since September 8, 2026, and Canada since September 18, 2026. Meta has not named a date for the UK, the EU, India, or Australia.

That matters because Muse is already an agent that acts, not a chat box. It can shop from a locked cloud computer with its own browser. It can also drive agent checkouts that left banks the bill when card networks treat a bot as a merchant risk. Wonder is the model Meta wants sitting on top of those flows. Until it actually does, Spark remains the brain in the app that people in the US and Canada already use.

Zuckerberg’s own wording left little doubt about the order of customers.

You can use Muse Wonder to power your Muse by subscribing today.

Mark Zuckerberg, CEO, on X

Specialized models “for everyone using Muse” sit after that. Builders on the API can buy Wonder now. People who only open the Muse app cannot.

CBRN Safeguards Come With a Report Still Pending

A personal agent that reads messages, calendars, and private documents needs discretion, Meta said, plus resistance to adversarial instructions. Wonder is described as the company’s most aligned model so far, with its strongest safeguards to date against chemical, biological, radiological, nuclear, and cyber misuse, while still exposing frontier cyber skills for defensive use under the Meta Superintelligence Scaling Framework.

WHAT WE KNOW

• Access today: Wonder is on the API and Muse Code; Full is limited to the $50 and $200 Muse Code plans.
• Data default: Muse Code subscriptions start in Contributor mode; API Wonder is Standard only.
• Safety paper: Meta says a full report with extra safety evaluations will be available soon, with no date.

WHAT IS UNCONFIRMED

• Full on the API: Docs do not say whether Wonder Full will ever get a public endpoint.
• Muse app timing: “Soon,” and subscribers first, is the whole schedule.
• Other countries: No Meta page names a UK, EU, or India date for the Muse app.

The launch Meta asked the world to celebrate is real. Independent tests put Full in the top tier for agent coding, 8 index points above Spark 1.3, at a 2.8x task cost for the API SKU. The people who can actually press that stronger model into a repo this month are the ones paying for Muse Code’s upper two plans. Everyone else is still on the wait Meta described as “over time.”