Muse Wonder Puts Its Strongest Model Behind Paid Plans

Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, a pair of agent models with a 1,048,576-token window. Only the first of them is on the public Meta Model API.

Muse Wonder Full, the stronger twin, stays inside Muse Code on the Power Usage and Maximum Usage plans. Independent tests still put Claude Opus 5.5 ahead on a broad intelligence index, while Meta’s own charts show Full pulling away on long coding jobs and legal agent work. The Muse personal agent, which Meta says these models will power, has no public date for the switch and will go to paying users first.

Two Models Open Two Different Doors

Meta Superintelligence Labs framed the drop as the next rung on a scaling ladder that started with Muse Spark on April 8, 2026. The research post says the new models were trained at the company’s gigawatt-scale Prometheus data center, after what Meta calls a 14-month ground-up overhaul. Mark Zuckerberg, Meta’s CEO, said the lab itself is just over a year old, and that Wonder is a much larger class of model than Spark.

Developer docs list the public model as muse-wonder-1.0, in preview. It takes text, image, video, audio, and PDF, returns text, and exposes reasoning effort from “minimal” to “max”. Turning that dial to “max” does not switch you onto Full. Meta’s model docs say Full is not available on the Model API.

WHERE EACH MODEL RUNS

Standard mode means prompts, files, and replies are not used to train Meta models. Contributor mode, which trades cheaper or extra usage for training rights, is off the table for Wonder on pay-as-you-go API traffic. Inside Muse Code, Contributor mode is on by default for subscribers, and it can be switched off in settings. That is the path to Full: a coding agent, on two named plans, with training rights sitting on unless you turn them off.

Independent Scores Still Trail Opus 5.5

Artificial Analysis, which scores models on its Intelligence Index v4.3, put Muse Wonder Full at 56 and Muse Wonder at 52 at max reasoning effort. Muse Spark 1.3, released about six weeks earlier, sat at 48. Claude Opus 5.5 leads that index at 58. Claude Sonnet 5.5 ties Full at 56. GPT-6 Astra is at 53.

The same lab said Full’s biggest jump over Spark 1.3 was Terminal-Bench 4.0, from 33% to 57%. On its Coding Agents Index, Muse Code with Full scored 68 at max effort, level with Claude Code running Sonnet 5.5 and ahead of Opus 5.5 at 66. Muse Wonder scored 63. Spark 1.3 on Muse Code scored 54.

Meta’s own launch chart, which the company published beside those third-party figures, is a closer fight with a different shape. Full leads on long software jobs and on Harvey’s legal-agent test. Opus 5.5 still wins a from-scratch coding test and the broad independent index.

SELECTED SCORES AT LAUNCH

Full’s SWE-Marathon number is the xhigh run Meta published. On DeepSWE 1.1 the two Wonder models are 0.2 apart, so the API model is already close on that long-coding set. Harvey LAB-AA is the gap that looks wide in print and small in life: 13.1 still means most legal work products fail a test that only counts a pass when every criterion is met and hallucination checks clear. Knowledge is the soft spot Artificial Analysis flagged. On AA-Omniscience, Full scored 30 against 46 for Opus 5.5. Full’s hallucination rate on that suite fell to 23% from 33% on Spark 1.3; Wonder is at 27%.

The Agent That Keeps Your Context

The product Meta is actually selling is not another chat box. The research post leans on Active Memory and Active Focus, two names for jobs most chat models still drop. Active Memory is meant to carry what matters about a person across conversations and pull those details back when a new task needs them. Active Focus is meant to keep working when a job gets hard, try another route, recover from errors, ask for a check when it should, and pause before an important action.

Alexandr Wang, Meta’s chief AI officer, listed three launch highlights: memory across days, weeks, and months; finishing tasks; and domain scores he called state of the art in finance and legal work, with a strong health showing. Meta’s demos show the intended use. One build turns custom images into a playable 3D kitchen game. Another works in a browser across apps to draft a Saturday run plan from personal preferences. Other clips have the model design swag from Instagram photos and sort a wedding seating chart from a knot of matching rules.

Those clips are also why the memory pitch matters more than the kitchen game. A personal agent that Meta says may handle messages, calendars, and private documents only works if it remembers the right slice of a life and then holds a goal long enough to finish it. Zuckerberg put the consumer door in one line.

You can use Muse Wonder to power your Muse by subscribing today. We’ll work to bring this level of intelligence to everyone over time.

Mark Zuckerberg, CEO, on X

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

What Muse Code Charges for Wonder Full

Muse Wonder on the API is billed per token at $4.00 input and $20.00 output per million tokens, with cached input at $0.20. Artificial Analysis priced that stack at $4.52 per Intelligence Index task, 2.8 times Muse Spark 1.3 at $1.60, mostly because the new rates are higher and Wonder reads more input. Full has no per-token price because it is not sold that way.

Meta’s Muse Code page lists four monthly plans. Full is named only on the top two.

MUSE CODE MONTHLY PLANS

• Everyday Usage: $5 a month, access to the Muse model family, and 10 to 50 prompts every 5 hours.
• High Usage: $15 a month, with 5 times the usage of Everyday Usage.
• Power Usage: $50 a month, 20 times Everyday usage, and access to the most powerful Muse models.
• Maximum Usage: $200 a month, listed for the most demanding coding work.

Launch day burned quota faster than users expected. Wang posted that evening that the team had identified and fixed unusually fast usage against limits for both Wonder models, and that Muse Code had been reset.

we have identified and fixed the issue that was causing the unusually fast quota usage for muse wonder and muse wonder full. you will now see much lower usage against your limits.

also, as promised, now that we've fixed the issue…

muse code reset!! pic.twitter.com/j7cldu9ybo

— Alexandr Wang (@alexandr_wang) October 12, 2026

That fix does not change the door. The model Meta is using to claim a frontier lead on several agent tests is a $50 or $200 Muse Code feature, and the API product is the cheaper, slightly weaker twin.

Safety Claims Rest on a Report Still Coming

A personal agent that can act across messages, calendars, and documents needs a different kind of restraint than a chatbot that only replies. Meta says it trained the Wonder family to refuse sharing private information without cause, to turn down help with harmful acts, and to stay useful for defensive security work such as finding flaws in a developer’s own code. It also says Wonder has its strongest safeguards yet against chemical, biological, radiological, nuclear, and cyber misuse, while still allowing frontier cyber skills for defense, under the Meta Superintelligence Scaling Framework.

The company calls Wonder its most aligned model to date. A fuller safety report with extra evaluations, it said, will be available soon. Until that paper is public, the privacy claim is a training goal attached to a product that is about to be pointed at inboxes and calendars, first for people who pay.

Overnight Tests Split From the Launch Charts

Coding is where Meta’s tables look strongest, and where public checks got messier within hours. Wang replied to a complaint that Wonder had not been trained like rivals on 3D demo games, and said the team had not explicitly trained on that genre. “Please try muse wonder on your actual coding work,” he wrote. The 3D kitchen clip in the launch video was always a demo of range, not the job the scores measure.

Public 3D prompts still landed as the first taste many people tried, and several of those runs looked thinner than rival clips on the same brief, including a lava-lamp scene that came back as a red cylinder with no glass and no lava. Dan Shipper, who runs the Every newsletter, posted a different check that is closer to the coding claim: a rewrite-from-scratch test on a sloppy codebase, where Muse Wonder scored 52% against 71% for GPT-Sol 6.1. That lines up with ProgramBench on Meta’s own chart, where Opus 5.5 is at 70.5 and Full is at 59.0. Long, test-graded repository work is the row Meta won. Building a program from a blank page is the row it did not.

Muse Gets the Model After Subscribers Do

Muse, Meta’s personal AI agent, launched on September 8, 2026, on Muse Spark. Wonder is scheduled to power more of that product, including artifact generation and heavier jobs, and Meta says it will later use Wonder to train smaller models for everyone on Muse. The newsroom post says the new models will roll out first to subscribers.

WHAT WE KNOW

• The split: Muse Wonder is on the Model API in Standard mode; Muse Wonder Full is only on Muse Code’s Power Usage and Maximum Usage plans.
• The window: Both models use a 1,048,576-token context window, the same size Spark 1.3 already had.
• The index: Artificial Analysis scores Full at 56 and Wonder at 52, with Opus 5.5 at 58.

WHAT IS UNCONFIRMED

• Muse timing: Meta has not published a date for Wonder inside the Muse app, only that subscribers go first.
• Full on the API: Docs do not say when, or whether, Full will be sold per token.
• Safety paper: Extra evaluations were promised as a follow-up, with no release day attached.

Zuckerberg said Meta will work to bring this level of intelligence to everyone over time. Muse Wonder Full remains a Power Usage and Maximum Usage feature, and the Muse app still has no public date for the swap.