Meta starts US rollout of Muse voice chats and animated avatars

Voice itself is already offered by competing AI services. Nor is delegating tasks exclusive to Muse: reporting at its launch identified Gemini Spark and Claude Cowork as other services that start tasks on consumers’ behalf. Muse’s animated presence is one part of a broader contest over useful agents.

A personable agent still needs permission

At launch, Meta said Muse could send emails, book travel, fill out forms and make purchases. Users choose which services to connect individually; where a service lacks an API, a software interface for connecting applications, Muse can use a browser instead.

That work runs in a dedicated virtual machine—a separate computing environment—in Meta’s cloud, with a browser visible to the user. A separate system called Sentinel controls access to the internet and connected services, deciding whether a proposed action proceeds, is blocked or needs user approval, according to the launch account. Meta says permissions can distinguish reading from writing and be limited by task, service, transaction or time. Meta says sensitive actions require approval, while previously authorized, lower-risk tasks can proceed.

The distinction matters: conversational warmth does not establish authority to spend money or change information. External services also retain control. The same account noted that restaurant-booking service Resy had said it would delete accounts using automated agents.

Meta’s privacy record makes its assurances consequential. The launch reporting recalled Facebook’s $5 billion settlement with the US Federal Trade Commission in 2019 over privacy violations. It also noted that Muse’s debut followed Meta’s August agreement to an $18 billion multistate settlement over social-media consumer harms. For Muse, Meta says conversations and data are kept out of its advertising systems, and that the agent cannot see users’ passwords or payment methods. That promise is separate from model data use: the launch account says queries to Meta models can be used by the company by default, with an option to turn that off. A confidential workspace that Meta itself could not see inside was still a planned release at launch, targeted before year-end.

What the performance numbers establish

Meta’s technical account describes Muse Realtime Voice and Muse Realtime Avatar as a streaming system: shared speech tokens drive both audio and generated video, keeping lips and expressions synchronized.

The company reports about 870 milliseconds from the end of a user’s turn to the first byte of synchronized voice and video, at 25 frames per second. It reports capacity for 12 concurrent video-generation sessions on a single GB200 hardware unit. These company figures concern responsiveness and video capacity, not the time or cost to complete a user’s task.

In comparisons with commercial avatar systems, Meta reports overall preference for Muse of 78% against Runway Characters and 88% against HeyGen LiveAvatar. Raters used each product’s native live calls for two- to three-minute conversations with matched avatar identities. This tests conversational presentation; the post gives no rater count, and the mannerism comparison with Runway was not statistically distinguishable from parity.

Independent observations are narrower. In a brief iOS test, the agent generally replied quickly but sometimes paused to look up or check information. Animation matched its speech well, although the tester found a customized human likeness uncanny. One short test cannot establish reliability across users or tasks.

Company-reported preference for Muse in separate live-avatar comparisons. These ratings assess conversational presentation, not task completion. Source: Meta AI Research. Source: Meta AI Research.

The next test is useful work at a sustainable cost

Muse launched with a free tier and Power and Maximum subscriptions costing $20 and $100 a month, respectively, for more usage. A payment card was required to get started. Meta said a meter would track remaining usage and users would be warned when free usage ran out. Meta said most users should fit within the free tier and that paid plans helped cover compute costs. At launch there was no advertising within Muse, while commerce opportunities were being explored. The voice announcement does not specify separate pricing, allowances or serving costs for the new feature, so the video-capacity figures cannot establish its unit economics.

Meta says voice conversations will come soon to its AI glasses and Muse Charm, a planned wearable with a small screen for the avatar. The rollout reporting says the device is expected before year-end; those extensions are still forthcoming.

The evidence needed next is task-level: how often spoken requests produce the intended result, under what permissions, and at what usage cost. Expressive conversation and favorable avatar ratings answer a different question from whether users should entrust more of their work and personal information to Muse.