한국어 English ✓

The Assistant You Talk To — and Build Apps With by Voice: Meta Muse Voice, Avatars and Spaces, and Why Korea Is Still Outside

As of Nov 15, 2026 · Checked directly against Meta's official newsroom, research posts and subscription help, and OpenAI's official help · Views -

For anyone who opens a personal-assistant app every day, Meta's first week of November reads in two steps. On November 5, real-time voice conversation and expressive avatars arrived in the Muse app; the next day brought spaces — little tools you build inside the app by describing them — and specialized models said to forget less. One blank runs through all three: nowhere in the official announcements does it say which plan gets how much of these features. Read against ChatGPT's voice-limits table — the most transparent pricing structure among rivals — here is what is confirmed and what is still empty.

The short version

Voice and avatars roll out in the US first — tap the phone icon, hundreds of voices, three avatar display modes, mid-sentence interruption (Meta newsroom, checked directly)

A space is a custom screen built by describing it — placement, scheduling, background auto-refresh; sharing is still "exploring"

Pricing is weekly tokens (Power $20/month for 500M a week, Maximum $100/month for 3B), but per-plan limits for voice and spaces, and any token conversion table, are unpublished

Muse voice and avatar research spec summary panel reconstructed from the official post
A panel reconstructed from the serving specs in Meta AI Research's official post. The 870ms figure and the rest are measurements of the research stack, and the post itself cautions that its examples do not all reflect the avatars available in the app. Source: Meta AI Research, "Bringing Your Muse to Life" (Sep 23, 2026)

Voice and a Face — 870 Milliseconds to the First Byte

Starting is simple. Tap the phone icon at the top right of the app and the conversation turns to voice. The intended use is not precisely engineered prompts but thinking out loud — starting a travel plan from a half-formed idea, or rehearsing an interview or a pitch aloud. Answers come back at once, and you can interrupt mid-sentence.

The face is optional. Keep voice only, put the avatar beside the conversation, or go full screen, face to face. The avatar shows expressions and gestures matched to the tone while Muse speaks, and reacts while you speak. Voices number in the hundreds and can be swapped at any time.

The technical spec lives in the research post from Connect day, September 23. Muse Realtime Voice produces a stream of voice tokens carrying both content and delivery; sound and avatar video read the same stream, which is how lips and expressions stay in sync. Video runs at 448×768 portrait, 25 frames per second, and latency was measured at about 870ms from the end of the user's turn to the first byte of the synchronized response. That number, though, is a measurement of the research stack, and the post itself cautions that its examples do not all reflect the avatars in the app. The face and speed you actually meet in the app sit outside what this spec guarantees.

The comparison against commercial avatar systems also has to be read as a company test. Against Runway Characters and HeyGen LiveAvatar, each in its own live-call experience, evaluators talked for two to three minutes and preferred Muse overall, the text reports. The detailed preference ratios live only in charts, and on mannerisms the Runway comparison was statistically indistinguishable from even — a caveat in the same post. Generated video carries an invisible, non-removable watermark (Video Seal) added with no latency cost.

Apps Built by Talking — What a Space Takes On

A space is a custom screen that appears inside Muse from a single description. Say a gift tracker, a daily word game, a DIY planning board, and Muse builds it; as you use it you can add what's missing or change its shape and behavior. Placement is chosen among the sidebar and main tabs, and a space can be hidden or scheduled — a weekly planner up front on weekdays, a weekend plan swapped in on Saturday.

The nine examples the company lists stick to daily routines: a shopping hub gathering Instagram saved wishlists and Marketplace messages, a school dashboard pulling children's events from school email into a calendar, a wedding planner bundling RSVPs and vendor invoices, a small-business order tracker, a freelancer's client hub. This list is a set of ideas the company proposes, not a list of things real users are already doing.

The space's real claim is automatic refresh. Connect the apps you use and new concert dates, price changes, and savings progress appear without being asked; checks, notes, and edits happen in place. Sharing does not exist yet — how to share a space with others is something the company says it is "exploring."

Underneath, the models changed too. Specialized versions trained with Muse Wonder's learning began rolling out to Muse and Meta AI, with two cores. Active Memory separates temporary from permanent information and prunes what has gone stale; Active Focus keeps a task moving when something unexpected blocks it. By the company's own measurement, forgetting key information needed for a goal fell 2.4-fold, and completion of everyday tasks with unexpected obstacles rose 67% — with the measurement design and baselines unpublished. Detailed chart values, whose computation Meta does not disclose either, are best read as directional company signals.

Muse versus ChatGPT pricing structure summary panel reconstructed from official help pages
A panel reconstructed from Meta's subscription help and OpenAI's Voice help. Muse publishes only a weekly token total, with nothing on how that token drains in voice; ChatGPT prints its voice limits in hours.

The Price Seat — an Assistant Counted in Tokens, Voice Counted in Hours

Muse's bill is written in tokens. On the official help page, Free carries a usage limit whose size is unpublished; Power, at $20 a month, allows 500M Muse tokens a week, and Maximum, at $100 a month, 3B a week. The problem is that no conversion table exists. Nowhere does it say how many voice minutes or space refreshes one Muse token trades for, nor whether the new voice, avatars, and spaces are included in the free limit or expand on paid plans — neither the announcements nor the help pages say. For now, the price of the new features can be written no further than "from $20 a month once you hit the limit."

Seat (official pages, checked directly)PriceVoice / assistant terms
Meta Muse Free$0/month, weekly limit size unpublishedPer-plan terms for voice, avatars, spaces unpublished. Voice rolls out in the US first
Meta Muse Power$20/month, 500M Muse tokens a weekNo token-to-work conversion table. Maximum is $100/month, 3B a week
ChatGPT Go$8/month (official pricing page listing, cross-checked)Live voice 3 hours per rolling 24h, smaller Quip model
ChatGPT Plus$20/month (official pricing page listing, cross-checked)Live voice 3 hours per rolling 24h, full GPT-Live-1.1. Desktop voice draws on the usage budget at $0.05/minute
ChatGPT Pro$100 / $200 per month (Voice help table, checked directly)Voice 15 hours ($100) / unlimited ($200, abuse guardrails)

What the table shows is not a price gap but a disclosure gap. At the same $20 seat, ChatGPT prints a 3-hour voice limit and even the names of its Korean voices (Haetsal, Hanul) in an official table; Muse prints only a weekly token total, with nothing on how that token drains during voice. Performance comparison is asymmetric too. Muse's 870ms is a research measurement to first byte, ChatGPT prints no latency figure in this table, and the avatar form itself is a differentiator ChatGPT's voice does not have. The only seat where numbers can yet rank the two is transparency of the plans.

A Scene You Can Hold — the Maker's Confession, and Nobody Else's Yet

Three days after launch, confirmed real-use signals are thin. The most concrete is Zuckerberg himself posting on November 6 that he built a BBQ game, an event tracker, and a home-control dashboard as spaces — the head of the company that made it, using his own product, which is not market validation. The company's nine space examples and its voice scenes are likewise all pictures the announcement imagines.

No named outside user handling real work through the voice avatar, no independent latency measurement, and no outside replication of the avatar preference test were confirmed. TechCrunch and The Verge each covered the video calls and voice mode on launch day, but what those reports confirm reaches only the launch facts and availability. So this article's case section separates "what the company says it can do" from "who is already using it." The latter is still one person — the company's own chief.

Read With Care — the US First, the Unpublished Limits, the Unwritten Data Terms

First, the blank on the map. Voice and avatars start in the US, and voice on AI Glasses and Muse Charm is only "coming soon," with no date. Spaces are said to open everywhere Muse is available, but cross-checked availability lists (the US, Canada, Japan, Singapore, Mexico) do not include Korea. The official help says only that Muse is in limited testing and not available in all regions, instead of listing countries. Korean voice support is also absent from the announcements — the seat where ChatGPT, with Korean voice names printed in its official table, splits away.

Second, the unpublished limits. The size of the free limit, the conversion between Muse tokens and real work, and the per-plan allocation for voice, avatars, and spaces are all empty. Whether a voice conversation cuts off mid-way at the limit, or a space's background refresh stops, cannot be known before use.

Third, the silence on data. Neither announcement says anything about recording retention for voice conversations, processing of avatar video data, or any relation to advertising. For features you entrust with your face and voice, the absence of that sentence is itself material for an adoption decision. The Video Seal watermark marks generated video for viewers; it is not a data term for the person who handed the data over.

The conclusion splits cleanly. Confirmed: on November 5–6 Meta officially brought real-time voice and expressive avatars (hundreds of voices, interruption, three display modes) and build-by-description spaces (placement, scheduling, background auto-refresh) to the Muse app; underneath sit research specs like 870ms to first byte and 60-fold distillation, a company measurement of forgetting reduced 2.4-fold, and a price structure of an unpublished free limit plus Power at $20 and Maximum at $100 in weekly tokens. Not yet confirmed: per-plan limits and token conversion for the new features, Korean support and a Korean launch date, voice-data retention and advertising terms, and any measurement from outside the company. The test starts at your own seat: where Muse is available, measure how many minutes of voice the free limit actually reaches and how many days a space keeps refreshing itself. In Korea, that measurement itself is still outside the invitation.

Sources

Puzz, in one line

Muse's new face and voice have been announced, but how many hours a month you may use that face appears in no table yet. The adoption decision starts not from the feature list but on the day a table with the limits in it appears, ribbit.