
Avatar 6 vs Muse Realtime Avatar: One You Rent, One You Live Inside
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 180 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 223 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 98 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 79 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 199 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 114 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
The comparison gets made because both are the best avatar either company has shown, and both are new enough that nobody has an opinion formed by experience. But Avatar 6, released by HeyGen on 17 November 2026, and Muse Realtime Avatar, previewed by Meta at Meta Connect on 23 September 2026 and rolling out in the US Muse app since 5 November, are not two answers to the same question. One is a renderer you call with audio and get a file back. The other is a face bolted onto a conversational agent inside a consumer app. The useful question is not which looks better. It is which one can exist anywhere near your product — and for one of these the answer is "not yet, and not on your terms." The part of this stack you can choose today is the text layer — one key for the script and routing layer and it is the piece that decides whether the avatar says anything worth rendering.
The shape of each system
Avatar 6 is stateless in the way that matters to an engineer. You send it a reference image or a short video, you send it the speech audio, and it returns a video file that runs the length of that audio. Nothing persists. There is no session, no memory, no conversation state. A twenty-six-minute render is one continuous generation from one audio track, and if you want a second take you send the request again.
Muse Realtime Avatar is the opposite: it has no meaning outside a session. It consumes speech tokens produced by Meta's conversational layer and turns them into synchronized face, lip motion and gesture as the conversation unfolds. Its published figure is 870 milliseconds from the end of your turn to the first byte of a synchronized voice-and-video reply — a first-byte number, not a time-to-complete-reply number. Its output is 448×768 portrait video at 25 frames per second, watermarked with a transparent overlay so a recipient can tell it is not a camera feed.
• Unit of work — Avatar 6: one request, one file, length set by the audio. Muse Realtime Avatar: one live session, no file, output only meaningful in the moment.
• Input — Avatar 6: reference media plus audio you generate or record. Muse Realtime Avatar: speech tokens from Muse Realtime Voice, a reference image, and a rolling window of recent video latents.
• Output — Avatar 6: 1080p or 2K video, up to 2880×1620, at whatever aspect the reference implies. Muse Realtime Avatar: 448×768 portrait at 25 fps, watermarked.
• Reach — Avatar 6: documented REST API, published in developer docs with two model identifiers. Muse Realtime Avatar: the Muse app, on 18-and-over accounts, in the US.
• Price — Avatar 6: subscription credits with no published per-second rate. Muse Realtime Avatar: no rate card of any kind; access comes with a Muse subscription.
• Evaluation — Avatar 6: vendor studies plus one independent arena placing it first. Muse Realtime Avatar: vendor-reported comparisons only.
• Control — Avatar 6: motion_prompt up to 1,200 characters, two-level expressiveness. Muse Realtime Avatar: no developer-facing controls at all.

Why the latency figures are not comparable
Meta's 870 milliseconds and HeyGen's silence on generation time are both worth reading carefully, because they measure different stages of different pipelines.
Meta's figure is the gap between you finishing a sentence and the first byte of a reply that contains both voice and video. It says nothing about the rest of the call — not sustained frame rate, not what happens when the topic changes mid-sentence, not whether the visual holds up after three minutes. Meta's own methodology compares against baselines Meta selected. It is a responsiveness claim, and a real one, but it is a single point on a curve.
Throughput is the axis Meta's number does not touch, because Meta is not selling throughput. The independent arena timed Avatar 6 at a typical 175.6 seconds per personal-avatar render — slowest of the five systems compared — with one render in ten running past nine and a half minutes; Runway's entry averaged 60.2 seconds over the same set. A batch pipeline absorbs that: queue a hundred scripts, collect them in the morning. A live call cannot, which is the constraint Meta optimised away and the reason the two systems are not interchangeable even where the picture quality overlaps.
So the two numbers do not compete. One is a real-time figure for a system that must answer inside a conversation. The other is an absence, for a system that was never built to answer inside one.
The evidence gap, stated plainly
The two systems are also unequal in how much a stranger is able to check. Avatar 6 has a public endpoint, so New Measure could enter it blind against seven rivals and publish Elo — first on both boards at 1,197 and 1,237 — alongside its own caveats about short scripts, 35-second clips and three competitors speaking through HeyGen's voice clone. Muse Realtime Avatar has no endpoint, so every comparison that exists was run by Meta, on a baseline Meta selected. That is not an accusation of dishonesty; it is what an app-only product can offer, and it is the same reason a developer cannot benchmark it either.

Muse Realtime Avatar has nothing equivalent, and cannot have. There is no endpoint for a third party to instrument and no public interface to script a comparison against. Meta reports preference splits against Runway Characters and HeyGen's LiveAvatar, and its own post discloses that the mannerism comparison against Runway was not statistically distinguishable from parity — a caveat worth more than the headline split it qualifies. Meta also notes that its demonstrations "do not all reflect avatars available in the Muse app," which is the sentence to keep in mind when watching any of them.
That asymmetry is the whole argument in miniature. A model with an API can be tested by strangers. A model that lives inside an app can only be tested by users, and users do not publish protocols.
What Muse Realtime Avatar is genuinely for
Strip out the comparison and Meta has built something with no close equivalent in HeyGen's line. Muse Realtime Avatar is the visual layer of an agent that already has memory, tools and a persistent cloud computer — Meta's Muse Secure VM is a standing Linux machine with a browser that the agent operates on your behalf. The avatar gives that agent a face during a video call, and the system perceives your video as well as your audio, reading gaze and expression rather than waiting for you to stop talking.
That combination — an agent with a durable workspace, a face, and the ability to see you — is a product category HeyGen is not in. HeyGen sells video production. Meta is selling presence. If your use case is "a patient-facing assistant that a person can look at while it works," Muse Realtime Avatar is the closest thing shipping, and the fact that it ships only inside Meta's app is a distribution constraint rather than a capability one. If your use case is "forty localised explainer videos by Friday," Meta has nothing for you.
The practical difference: consent and control
Putting a real face on Avatar 6 means clearing a gate with no analogue in a consumer app. A reference video of an actual person needs Enterprise access, an on-camera consent statement bearing a per-session code, and a match check between the person consenting and the person in the footage; skip any of it and the request dies with a 422 rather than degrading quietly into something shippable. Meta handles the same problem by policy — 18-and-over accounts, US only, and a visible watermark on every frame it produces.
Meta's constraints are distributional instead. The Muse app is 18-and-over, the rollout has been US-only since 5 November, and developer access is not on offer — Meta's developer catalogue lists Muse Code, Muse Wonder, Muse Spark, Muse Spark Open, Muse Glimmer, Muse Image and Muse Voice Transcribe, and neither Realtime product appears there. You are not integrating Muse Realtime Avatar. You are, at most, using it.
What OrcaRouter has to do with either
Avatar 6 is not something OrcaRouter can route: it is a HeyGen render, billed in HeyGen credits behind HeyGen's consent flow, and Muse Realtime Avatar has no endpoint at all. What we do serve is the Muse model underneath Meta's agent stack — meta/muse-spark-1.2, at provider list price with zero markup — plus the text models that write the scripts Avatar 6 performs.
That is the honest boundary, and it is where the practical advice lives. If you are building a conversational product with a face, the reasoning and tool-calling layer is the part you can choose today, and it is the part that determines whether the thing is useful. The renderer is one vendor decision on top. Putting that layer behind a single key with automatic failover means a provider outage costs you a retry rather than a product.

Who should care about which
If you produce video — training, marketing, localisation, outreach — Avatar 6 is the more capable tool by a wide margin, and the only one of the two you can put in a pipeline this month. Its long-form capability is unevaluated by anyone outside HeyGen, its per-second price is unpublished, and the arena that ranks it first never tested a clip longer than 35 seconds. Go in with those three caveats and it is still the strongest option available.
If you are building an agent that people will talk to face to face, Muse Realtime Avatar is the more interesting system and the less usable one. Watch it. Do not architect around it. Meta has shipped it to consumers first and developers not at all, and until there is an endpoint there is nothing to build against.
What would change this page: a Muse developer API, or an independent real-time evaluation of Muse Realtime Avatar. Either would let the two be compared on something other than intent.
