We’re sharing early results from Avatar Arena, based on blind votes from creators comparing eight avatar providers. In each comparison, two avatars read the same script, and the creator votes before seeing which provider made either video.
HeyGen Avatar 6 was preferred most often on both personal and stock avatars, and Creatify Aurora 2.0 ranked second on both. Below we cover how each provider fared, where providers came closest, how long videos took, and the limits of what this round can tell us.
What we compared
A creator writes a short script. Two providers each make a video of an avatar saying those words. The videos appear as A and B, with their order chosen at random and the provider names hidden. The creator watches both and chooses a winner, says they are equally good, or says both are bad.
We compare two kinds of avatar, each with its own leaderboard:
- Personal avatars are AI versions of the creator. They upload a video of themselves, and both providers use that upload to make a new video of them reading the script.
- Stock avatars are ready-made presenters. Each provider supplies its own presenter and voice, so the two videos feature different people reading the same script.
After voting, creators can also answer questions about what they liked: which avatar moved more naturally, looked more real and matched its mouth movements to the speech. For personal avatars, we also ask which looked and sounded more like them. The provider names are revealed after these questions.
Overall results
Personal avatars
Stock avatars
Higher scores mean an avatar was preferred more often. Each leaderboard has its own average score of 1,000, so personal and stock scores should be read separately.
HeyGen Avatar 6 scores highest in both categories, by more than the uncertainty in the scores. Creatify Aurora 2.0 is second in both. On personal avatars its range overlaps with VEED Lipsync 2.0, so we can’t yet separate the two. On stock avatars, Creatify is clearly ahead of the providers below it. The votes don’t give us enough evidence to rank Tavus, JoggAI and Synthesia against one another, or Captions against Runway.
Where providers came closest
After voting, people could also say which video was better at specific things, like lip sync, voice or movement. They could also say both were about the same.
- Creatify Aurora 2.0 matched or beat HeyGen on lip sync (how well the mouth fits the words) about one time in three, for both personal and stock avatars. On personal avatars, it did best on voice: 38% of answers rated it as good as or better than HeyGen. But Creatify’s personal-avatar voice was made with HeyGen’s voice cloning, so this result partly reflects HeyGen’s voice (see How we counted).
- Tavus Phoenix-4.5 and JoggAI matched or beat HeyGen on lip sync about one time in four.
- Movement showed the biggest gap. Against Runway, Captions and Synthesia, at least nine out of ten answers said HeyGen’s avatar moved more naturally.
Speed and reliability
Speed and preference didn’t line up. Runway GWM-1 Avatars was the fastest for personal avatars, typically about one minute per video, and JoggAI the fastest for stock presenters, at about 46 seconds. HeyGen Avatar 6 typically took about three minutes, Creatify Aurora 2.0 a little over two, and Captions Avatar X about eight minutes.
Typical times hide the slow cases. One in ten HeyGen personal-avatar videos took more than nine and a half minutes, compared with about three minutes for Creatify and under a minute and a half for Runway. For stock presenters, HeyGen and Tavus had the slowest cases, with one in ten videos taking more than seven minutes.
Some videos failed to generate: about 7% for Captions, 3–4% for Runway and Synthesia, and about 2% or less for the other providers. A failure can come from our integration as well as from the provider.
People need both videos before they can compare them. From starting a comparison, the usual wait was about five minutes for personal avatars and four minutes for ready-made presenters.
Some waits were much longer. About one in ten personal comparisons took more than 17 minutes. For ready-made presenters, about one in ten took more than 13 minutes.
What about video length?
The typical script was 62 words. In 97% of completed comparisons, both videos were 35 seconds or shorter.
We limit scripts to 500 characters, so people could only make short videos. These results don’t tell us which provider would do best at a two-minute explanation or a ten-minute lesson. They also don’t tell us how long people would want their videos to be without that limit.
What people still wanted changed
After voting, people could ask for changes to the video they chose, using an editor in the arena. About one in six voters (17%) used it.
Even among people who preferred the HeyGen video, many requests were about how the avatar looked, moved or sounded, not just what it said. One creator wrote: “Last line: when I'm saying that we have to shut it down, I'm smiling. Can you make it a little serious?” Others wanted fewer hand gestures, a more natural voice, a different outfit or a different background.
The editor can change the words and generate another video. It has no controls for hands, clothes, backgrounds or facial expressions, so asking it to fix those things can produce another video without fixing the problem.
That is worth testing across all providers in a future round: can an avatar follow a specific instruction, such as keeping the words and voice the same while lowering its hands or making its expression more serious?
How we counted
These are early results from people who chose to use Avatar Arena. Their votes may not represent everyone who makes avatar videos.
- Dates and accounts. We count votes and edits from 27 October through 16 November 2026.
- Scores. We use a ranking method called Bradley–Terry and display the results on an Elo scale, with an average of 1,000 in each category. A tie or “both bad” contributes half a point to each side.
- Robustness. Counting only each person’s first vote in each category, HeyGen stays first and Creatify second in both. The personal-avatar order is unchanged; the stock order below Creatify changes, consistent with those providers’ overlapping ranges.
- Failed videos. In about 3% of votes, a video had failed to generate, most often the other provider’s, and people voted on what they saw. We counted these as normal votes. Leaving them out changes no ranks and moves scores by at most 11 points.
- Generation times. We start timing when a provider accepts a request to make a video and stop when the arena saves the finished video. This excludes avatar setup and speech preparation.
- Provider setups. These results cover the versions and settings we tested. For personal avatars, HeyGen and Runway create their own avatars from the upload. VEED changes the mouth movements in the uploaded footage; Creatify and Captions animate a still image from it. Those three use speech made with HeyGen’s voice clone, so their voice ratings partly reflect HeyGen’s voice technology. Animating a still image may also limit movement compared with using the full video. We haven’t tested enough languages or long videos to draw conclusions about those uses.
