Model

HeyGen's Avatar 6 scored 84% on realism, but its 42.2% pick rate was below chance

Avatar 6 launched November 17, 2026, with an 84% realism figure and first place in a blind arena; the score doubles a below-chance pick rate, and the arena capped 97% of clips at 35 seconds. That is the evidence base under a company claiming $200 million of annual revenue.

Vincent JiangVincent Jiangthree-minute read
Share
Joshua Xu, chief executive and co-founder of HeyGen, speaking on stage at Collision 2024
Joshua Xu, chief executive and co-founder of HeyGen, on the PandaConf stage at Collision 2024 in Toronto, where he argued that viewers decide in seconds whether a video feels like a person talking to them.

The 84 percent, taken apart

HeyGen~$500M — HeyGen, private, latest valuation ~$500M launched Avatar 6 on November 17, 2026, an audio-driven avatar model that reads a script and times the gestures to the words 1. The headline evidence: 84% of participants in a Turing-style test found the avatar as realistic as the person it represents 1. Chief executive Joshua Xu sells trust, and names the stake: "People decide in seconds whether a video feels like a person talking to them, and when it doesn't, they stop listening" 1.

Of 270 side-by-side answers, 42.2% picked the avatar

The technical report published the same day scores indistinguishability at 84.4% and shows the arithmetic underneath: outside raters chose Avatar 6 as the real recording in 42.2% of 270 side-by-side answers, below the 50% a coin flip delivers, and the headline score doubles that share 2. A separate single-clip test had 65.1% of answers calling Avatar 6 real, against 78.9% for genuine recordings 2. Doubling a below-chance pick rate does not move it above chance.

Raters judged the real person real more often than the avatar

The avatar scored below the real recording, and below a coin flip head-to-head

0%20%40%60%80%Genuine recording78.9%Avatar 6, own voice (single clip)65.1%Avatar 6 (head-to-head pick)42.2%Coin flip
Data
Value
Genuine recording78.9%
Avatar 6, own voice (single clip)65.1%
Avatar 6 (head-to-head pick)42.2%
Share of answers judging a clip real in HeyGen's external-rater tests, November 2026. The report's 84.4% indistinguishability headline doubles the 42.2% head-to-head pick rate. Source: HeyGen technical report via Unite.AI [2].2

First place, measured on half-minute clips

The independent evidence is real but short. New Measure's Avatar Arena, blind votes across eight avatar providers from October 27 to November 16, ranks Avatar 6 first in both categories, Elo 1,197 for personal avatars and 1,237 for stock presenters against a 1,000 average 3.

The same results page draws its own fence: the typical script ran 62 words, 97% of comparisons had both videos at 35 seconds or shorter, and the round cannot say which provider wins a two-minute explanation, let alone a long lesson 3. HeyGen's demonstrations run 25 to 28 minutes 2. Renders typically took about three minutes against roughly one for Runway's GWM-1 avatars, and one in ten HeyGen videos ran past 9.5 minutes 3. Even the runner-up ran partly on HeyGen: Creatify's personal-avatar voice used HeyGen's voice clone 3.

The number that survives, on HeyGen's own metric

One result holds up on any reading: HeyGen's editor regenerates only the edited span, changing 34.4% of frames on average across 20 edits, while MiniMax's H3-max, Alibaba's−2.51% — Alibaba, down 2.51 percent today Wan 3.0, ByteDance's Seedance 2.5 and HappyHorse Edit rewrote 99.7% to 100%, the whole take 2. Against general video models the edge narrows: 50.1% versus Wan 3.0 from a photo, a tie by HeyGen's own label, 54.9% versus Kuaishou's Kling 4.0 and 72.4% versus Seedance 2.5 2.

HeyGen's editor rewrites a third of the frames; four rivals rewrite nearly all

0%20%40%60%80%100%MiniMax H3-max100%Alibaba Wan 3.0100%HappyHorse Edit100%ByteDance Seedance 2.599.7%HeyGen Avatar 6 editor34.4%
Data
Value
MiniMax H3-max100%
Alibaba Wan 3.0100%
HappyHorse Edit100%
ByteDance Seedance 2.599.7%
HeyGen Avatar 6 editor34.4%
Share of frames rewritten after 20 edits, HeyGen's own external-rater test, November 2026. HeyGen's editor regenerates only the edited span; the four competitors rewrote 99.7% to 100% of the take. Source: HeyGen technical report via Unite.AI [2].2

Against general video models, Avatar 6's edge runs from a tie to 72%

HeyGen tied Wan 3.0 and edged Kling 4.0; the clear win is Seedance 2.5

0%20%40%60%80%vs Seedance 2.5 (ByteDance)vs Kling 4.0 (Kuaishou)vs Wan 3.0 (Alibaba)Even split50.1%
Data
Avatar 6 share of preferring answers
vs Seedance 2.5 (ByteDance)72.4%
vs Kling 4.0 (Kuaishou)54.9%
vs Wan 3.0 (Alibaba)50.1%
Share of head-to-head answers won by Avatar 6 among raters expressing a preference, HeyGen's own external-rater tests, November 2026. Wan 3.0 compared from a photo; Kling 4.0 and Seedance 2.5 from footage. Source: HeyGen technical report via Unite.AI [2].2

The private mark, and the listed labs it benchmarks

HeyGen is private, last priced at $500 million in June 2024 on $35 million of annual recurring revenue, and says it passed $200 million in June 2026, doubled in eight months, an unaudited run rate 1 4.

Public money sits in its comparison table: MiniMax, which floated in Hong Kong in January and opened H3's weights within days of release, is now graded on an edit metric written by a company that shipped on ByteDance's Seedance 2.0 in April and now sells its own avatar, video and voice stack 5 6. The consent story has a seam, too: the release requires explicit consent for realistic videos of real people, while the API documentation restricts realistic video references to enterprises and lets some contracts skip the consent step 1 7.

Watch the long-form round

The open question is duration. Until a long-form round of the arena runs, a product demonstrated at 28 minutes is being marked on evidence gathered 35 seconds at a time 3.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading BABA

Reader comments

0 comments

    Model

    MiniMax hit a record low, about 90% below its January peak, as a US rule gates its rented Nvidia compute

    Washington's new remote-access rule puts a US license between China's listed model makers and the offshore Nvidia capacity they rent. MiniMax, the purest listed bet on Chinese foundation models, holds the price tag.

    Model

    OpenAI cuts flagship pricing 20% to match Anthropic as heavy users call GPT-6.2 Astra unusable

    OpenAI cut its flagship's price to match Claude Fable 5.2's rate, and the users who run models hardest rejected it, 24 hours before Anthropic priced a $1.4 trillion IPO on the assumption the quality gap holds. The independent labs say the gap is real, for now.

    Model

    Microsoft books Azure revenue on Chinese open models running on Baseten hardware it does not own

    Microsoft now bills customers' committed Azure dollars for open models from Chinese labs running on Baseten's hardware, a toll booth earning rent on compute it does not own.

    Model

    Alibaba stops giving away its best AI model, three days before earnings

    Alibaba launched Qwen4-Max on November 16 and kept the 2.4-trillion-parameter flagship behind its own paid API, promising open weights only for smaller models. Thursday's report is the first test of whether Qwen's three billion downloads become cloud revenue, and it lands two days after a US rule put the chips behind the model under license.

    Model

    Perplexity claims $1 billion of revenue on five times less spend than Anthropic's $1.4 trillion IPO

    Perplexity says it hit $1 billion in recurring revenue on a fraction of Anthropic's spend, three days before Anthropic prices a $1.4 trillion IPO on the opposite assumption. Whose economics hold will only be settled by a filing Srinivas now says could come in early 2027.

    Model

    Anthropic prices a $1.4 trillion IPO as Databricks logs show its costliest model at 0.6% of enterprise tokens

    Databricks' gateway logs show the costliest Claude tier at 0.6% of enterprise tokens and open-weight models near a third, hours before Anthropic prices a $1.4 trillion debut on premium-model revenue.

    Subscribe to TheInference for free