Muse Realtime Avatar vs GPT-Live-1: Presence Against Conversation

The billing unit tells you what each company thinks it is selling. GPT-Live-1, Open​AI's full-duplex voice model, bills at a flat $0.05 per minute of voice charged by the second, and its rate limits are denominated in concurrent sessions — 25 on Tier 1, rising to 500 at Tier 5. That is a phone line's unit. Muse Realtime Avatar, announced by Meta on September 23, 2026, has no billing unit, because there is nothing to bill: no API, no endpoint, no price, no date. Its published unit is a hardware figure instead — 12 concurrent real-time sessions on one NVIDIA GB200. One company is selling conversation by the minute. The other is showing you what a face costs to render, and has not decided to sell it yet.

Both models are new-ish and neither is a launch in the sense the word usually carries. GPT-Live-1 shipped inside ChatGPT on July 8, 2026 and only reached the developer API on September 10 — two months during which it was the default voice of a consumer product and unavailable to anyone building. Muse Realtime Avatar was announced September 23, 2026 at Meta Connect, is demonstrated in Meta's own research post, and remains a capability of the Muse agent rather than a product. Comparing them is comparing two different stages of the same idea: what happens when a conversational model is good enough that the next question is what the user is looking at while it talks.

What OpenAI built: a conversation that never stalls

GPT-Live-1 is full-duplex in the sense that matters operationally — it is not waiting for you to finish before it starts working. It reaches the developer API through exactly one endpoint, the Live Sessions endpoint, under exactly one model ID, gpt-live-1, with no dated snapshots to pin and no sibling endpoints enabled. No Chat Completions, no Responses, no Realtime, no fine-tuning, no audio transcription or speech endpoints. Image and video input are explicitly unsupported.

The design decision that shapes everything downstream is delegation. GPT-Live-1 does not do its own hard thinking. Open​AI's documentation pairs the voice layer with a separate reasoning backend, and in the published WebRTC example that backend is GPT-5.6 Terra. So a GPT-Live-1 session is two meters running at once: $0.05 a minute for the voice, plus the backend model's normal token rates for every tool call, search and reasoning step it hands off. On the Speech to Speech Index that split personality shows up as the same model scoring twice — 81.5 with GPT-6 Astra doing the thinking at medium effort, 80.1 with GPT-5.6 Sol at low effort. The 1.4-point gap between those two configurations is wider than the 0.2-point margin that puts GPT-Live-1's best configuration in first place.

That is the honest shape of the model. The voice front end is fixed; the intelligence is a parameter you set, and it moves the score more than any competing model does.

What Meta built: a face that moves with the voice

Muse Realtime Avatar attacks a problem GPT-Live-1 does not have — keeping generated video in lockstep with generated speech. Meta's answer is to make them the same stream: Muse Realtime Voice emits speech tokens carrying both content and prosody, and the avatar is an audio-driven Diffusion Transformer consuming those same tokens, conditioned on a reference image and a rolling window of recent video latents. Nothing is lip-synced after the fact, because there is no "after the fact" — voice, mouth and expression come out of one process.

The published figures are Meta's own, measured against baselines Meta chose, and none has been reproduced outside the company. 870 ms from the end of your turn to the first byte of a synchronized voice-and-video reply. 448×768 portrait video at 25 frames per second, watermarked with a transparent overlay so a viewer can tell it is not a camera. Twelve concurrent sessions on one GB200, an 8× capacity gain over Meta's own two-step BF16 baseline, from four-bit quantization-aware training, persistent KV caches and latency-aware dynamic batching. And a preference result Meta reports going from 45% to 55% over that baseline, with two disclosed wins against commercial avatar systems — plus the disclosure that on mannerism, the result against one of them is not statistically distinguishable from parity.

Nothing in that list is a rate, an endpoint, or a date.

The two-meter bill, and where routing earns its place

GPT-Live-1's cost structure is not one number, and treating it as one is how a cheap-sounding voice model produces an expensive month. Voice is $0.05 a minute, charged by the second with no round-up, and the first 15 seconds of a session are billed up front but credited against later duration rather than stacked on top. Everything the session delegates — the reasoning, the tool calls, any search — bills separately at the backend model's own rates. Point delegation at a frontier reasoner and let it search on every turn and you have built a $3-an-hour open line with a premium model behind it.

That second meter is the routable half. On OrcaRouter the backend of a GPT-Live-1 session is just another model call: 196 models from 15 providers behind a single key, at provider list price passed through with zero markup, with automatic failover when the model you chose errors or times out. You cannot route the voice layer — Live Sessions is Open​AI's endpoint alone — but everything the voice layer delegates to sits on one key, which means the part of the bill that scales with how hard your callers think is also the part you can change without a contract.

Muse Realtime Avatar, by contrast, has no meters to route. It is a feature of an app.

Two bets about what a voice agent is for

Strip the specs and the two companies disagree about the product. Open​AI's bet is that the conversation is the product — that a voice agent is a phone line, and the work is making it feel like one: never stalling, handing hard thinking to a model behind it, billing by the minute, scaling by the concurrent call. Every choice in GPT-Live-1's API surface follows from that. Concurrency-based rate limits, WebRTC-only transport, a deliberately narrow single endpoint, no video in or out.

Meta's bet is that presence is the product — that a user who can see the agent is a user who stays in the conversation, and that the interesting engineering is not turn-taking but keeping a rendered character coherent for the length of a call. Those choices produce a system optimized for frame rate and session density on a GPU, with no commercial surface at all.

There is a real possibility both are right and the two get composed. A product could run its conversation on a full-duplex voice model and its visual layer on an avatar renderer, and the only thing stopping that today is that the avatar renderer has no endpoint. What is not plausible is treating them as two versions of the same purchase.

Who should act, and who should wait

If you are building a voice product now, GPT-Live-1 is callable and Muse Realtime Avatar is not, and that ends the decision. The interesting choice inside GPT-Live-1 is the backend you delegate to, because that is where both the score and the bill actually move — and it is the one part of the stack you can route, swap and fail over without touching the voice integration.

If what you want is an avatar, waiting is not passive. The things worth watching are specific: whether Meta publishes a rate card and an endpoint, whether a stated date appears, and whether 870 ms survives contact with a phone on a cellular connection rather than a Connect demo. Meta's own post notes that its demos do not all reflect avatars available in the Muse app, which is the sentence to hold onto.

Until one of those changes, the comparison has a one-line answer. GPT-Live-1 is a phone line with a brain you choose. Muse Realtime Avatar is a face with no phone number.