Meta Launches Muse Wonder, Its First Frontier Model, at $4 per Million Tokens While Muse Wonder Full Stays Subscriber Only

Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, calling them its most capable models and the first time its Superintelligence Labs has pushed into frontier territory. Muse Wonder is live today on the Meta Model API and in Muse Code. Muse Wonder Full, the stronger variant, is not on the API at all.

The API price is USD 4 per million input tokens and USD 20 per million output tokens, according to Meta’s pricing page. That puts Meta in the same bracket as the top models from OpenAI and Anthropic, and it means the benchmark claims need a closer look.

Mark Zuckerberg announced it on X with a post that had passed 3.7 million impressions by Tuesday morning IST, describing Muse Wonder as “our new family of frontier models with state of the art performance for personal and professional agentic work.”

What Meta actually shipped

Meta’s research blog post pitches the series around agents that work for you over long stretches. Two named capabilities carry most of the story: Active Memory, which carries context about you across conversations, and Active Focus, which keeps the model working through blockers on a long task.

The model card lists text, image, video and audio as inputs, text as the only output, and a 1M token context window at inference. Meta’s developer docs put the exact window at 1,048,576 tokens and add PDF as an input type, with reasoning effort settings from “minimal” to “max”.

Meta says both models were trained at its gigawatt-scale Prometheus data center. TechCrunch’s Russell Brandom reports that the site is in central Ohio and came online earlier this year.

Muse Wonder vs Muse Wonder Full: who gets what

This is the part most launch headlines skip. The model topping Meta’s own charts is mostly Muse Wonder Full, and you cannot call it from your own code.

Item	Muse Wonder	Muse Wonder Full
Meta Model API	Yes, as muse-wonder-1.0 (Standard mode only)	Not available
Muse Code	Pay-as-you-go and every subscription plan	Power Usage and Maximum Usage plans only
API price (USD per 1M tokens)	4.00 input, 20.00 output, 0.20 cached input	No API price
Muse consumer app	“Soon”, subscribers first	Not stated
Rate limits	3,000 requests and 4M tokens per minute, per team	Plan usage limits

Meta’s models page is explicit that setting reasoning effort to “max” on Muse Wonder does not select Full. Web search grounding costs an extra USD 2.50 per 1,000 queries, and Meta says there is no long-context premium.

The benchmark numbers, and where Meta does not win

Meta’s evaluation table compares both Wonder models at max reasoning effort against Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5. Here is a subset, as published by Meta.

Benchmark	Wonder Full	Wonder (API)	Opus 5.5	GPT-6 Astra	Fable 5.1
Personal Superintelligence Index (Meta internal)	59.1	54.3	55.2	54.5	49.9
MemoryBench (Meta internal)	85.5	85.6	81.5	81.4	79.0
Harvey LAB-AA (legal)	13.1	10.0	4.2	8.6	6.4
SWE-Marathon	57.5 (xhigh)	53.1	46.3	42.5	45.6
Terminal-Bench 4.0-AA	57.1	52.0	59.6	59.1	52.0
ProgramBench	59.0	50.0	70.5	52.5	55.5
Artificial Analysis Index	56	52	58	53	53

Read the table closely and the story is more mixed than “Meta caught up.” On Meta’s own sheet, Claude Opus 5.5 leads the Artificial Analysis Index, Terminal-Bench and ProgramBench, where it scores 70.5 against 50.0 for the API version of Muse Wonder.

The two headline wins, the Personal Superintelligence Index and MemoryBench, are internal Meta benchmarks. The model card also says that where no third-party result exists, Meta used “the highest of our internally validated run, self-reported score, or official leaderboard score,” and that competitor runs “may not be optimally tuned for proprietary models.”

Some rows do come from outside graders, including Artificial Analysis, Vals AI, Datacurve and Abundant AI. The SWE-Marathon row is not a common-harness comparison, though: GPT-6 Astra and Fable were scored through Codex and Claude Code. The clearest third-party win is Harvey LAB-AA, where both Wonder models beat every rival listed, though absolute scores on that test are low across the board.

Price check: same bill as Opus 5.5

VentureBeat’s comparison lists Claude Opus 5.5 at the same USD 4 and USD 20 per million tokens, and GPT-6 Astra at USD 10 and USD 50. So Muse Wonder undercuts Astra, but matches Opus 5.5 on price while trailing it on the Artificial Analysis Index in Meta’s own table (52 vs 58).

Meta’s cheaper Muse Spark 1.3 stays at USD 1.25 input and USD 4.25 output. Unlike Spark, Muse Wonder has no discounted Contributor mode on the API.

Muse Code subscriptions come in four tiers (Everyday, High, Power and Maximum Usage), but Meta’s subscription docs do not publish USD monthly prices. Meta says benefits may vary by region and are shown during onboarding.

Safety claims, with caveats Meta itself flags

Meta calls Muse Wonder its “most aligned model to date” and says it has its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse. The full safety report is promised “soon,” so outside reviewers have only the model card for now.

The card is frank in places. Without mitigations, Meta says it does “not rule out” Muse Wonder Full reaching its high risk threshold in the chemical and biological domain, so it added safety post-training and system-level classifiers. With those in place, Meta rates residual risk as “moderate or lower” in every domain.

One cyber detail is worth knowing. On ExploitGym, Muse Wonder Full captured 92 flags, but a verification pass found only 21 used the intended vulnerability, and 21 is the figure that matches the public ranking. Meta also lists 3D and spatial reasoning as a known weakness and says the model is not meant for users under 18.

Confirmed vs unconfirmed

Claim	Status	Source
Release date October 12, 2026; two models	Confirmed	Meta blog and model card
API price USD 4 / USD 20 per 1M tokens, USD 0.20 cached	Confirmed	Meta pricing page
Wonder Full is Muse Code only, Power and Maximum plans	Confirmed	Meta developer docs
1M token context; text, image, video, audio input	Confirmed	Model card
State of the art on personal and professional agent tasks	Vendor claim, partly internal benchmarks	Meta evaluation table
“Caught up” with OpenAI and Anthropic	Not independently verified; Opus 5.5 leads several rows	Meta’s own table
Muse Code subscription USD prices	Not published in docs	Meta subscription docs
Full safety report	Promised “soon”, no date	Meta blog
Internal codename “Watermelon”	Unconfirmed, secondary reports only	Not in Meta materials
Parameter count and architecture	Not disclosed	None

What it means for Indian developers

Muse Wonder is not on the India restriction list in Meta’s Geographic Use Policy, updated October 12. The only model blocked for India there is Muse Image, so check that before you plan an image feature on the same API key.

Cost math is simple. A job that reads 200,000 tokens and writes 20,000 costs about USD 1.20 on Muse Wonder, against about USD 0.34 on Muse Spark 1.3. For classification and short tool loops, Spark is still the sensible default, and Meta’s docs say so too.

Watch the data setting if you work on client code. Muse Code subscriptions default to Contributor mode, which lets Meta train on your prompts, code and outputs. API calls to Muse Wonder run in Standard mode, where Meta says your content is not used for training.

The API speaks OpenAI and Anthropic request formats, so trying it inside an existing agent is mostly a base URL and model ID change. If you are building on open agent standards, our look at the Personal Agent Protocol draft Meta backs with Sierra shows where Meta wants these agents to plug in.

The bigger picture for Meta

Fourteen months ago Meta began what it calls a ground-up overhaul of its AI work. Muse Spark came first, including the research run we covered when Meta said Muse Spark helped answer five open math questions. Muse Wonder is the first time Meta has put a model in the top price bracket and asked to be judged against the leaders.

The fair verdict today: a credible, competitive model with real third-party wins in legal and long-horizon coding, priced like a premium product, with its best version held back from developers. Independent tests over the next few weeks, like the ones behind Scale AI’s everyday visual benchmark, will matter more than launch-day charts.

FAQ

What is Meta Muse Wonder?

It is Meta’s frontier model family, released October 12, 2026, built for long-running personal and professional agent tasks and coding. It comes in two variants, Muse Wonder and Muse Wonder Full.

How much does the Muse Wonder API cost?

USD 4 per million input tokens, USD 20 per million output tokens and USD 0.20 per million cached input tokens, with web search at USD 2.50 per 1,000 queries.

Can I use Muse Wonder Full through the API?

No. Meta offers Muse Wonder Full only inside Muse Code on the Power Usage and Maximum Usage subscription plans.

Is Muse Wonder available in India?

Meta’s Geographic Use Policy does not restrict Muse Wonder in India. It restricts only Muse Image there. Availability can change, so check the policy before you ship.

Are Meta’s benchmark results independently verified?

Partly. Several rows come from third parties such as Artificial Analysis and Vals AI, but the headline Personal Superintelligence Index and MemoryBench are internal Meta tests, and the full safety report is not out yet.