Humans vote,
AI gets ranked

Some things only people can judge, like whether a voice sounds natural or an avatar feels real. Watch two anonymous outputs from the same prompt and pick the better one. The models are revealed after you vote.

Votes cast
2,000+
Models benchmarked
25+
Arenas
3
  • HyperFrames+ Opus 5.5

    1060plus or minus 46

    Wins 30–60 second films

  • Remotion+ Opus 5.5

    1053plus or minus 40

    Wins films over a minute

  • Remotion+ Sonnet 5.5

    1026plus or minus 48

    Top entry on Sonnet 5.5

Avatar Arena Leaderboard: Personal

view the leaderboard →
RankProviderelo95% CIIntervalVotes
1HeyGenAvatar 61,1971162 to 1241plus or minus 40542
2CreatifyAurora 2.01,0531008 to 1111plus or minus 52133
3VEEDLipsync 2.01,000943 to 1051plus or minus 54127
4CaptionsAvatar X896829 to 955plus or minus 63163
5RunwayGWM-1 Avatars854743 to 935plus or minus 96119

Personal track: each person uploads a video of themselves, and both providers use that footage to make a new video of them reading the same script. Based on 542 blind votes, 27 Oct – 16 Nov 2026. Bradley–Terry scores on an Elo scale, where 1000 is the average for this track, with 95% ranges from 1,000 bootstrap resamples of voters. A tie or “both bad” vote contributes half a point to each side. Every rival faces HeyGen; comparisons between rivals are indirect. The report assesses uncertainty in score differences. Data through 16 Nov 2026, excluding known internal accounts.

Methodology

How a vote becomes a rank

Blind by default

You see two outputs, never two names. Models are revealed only after you vote, so brand can't sway the result.

Same conditions

Both sides get the same prompt, inputs and budget. We test the system as it ships: model, harness and settings together.

Real prompts

Prompts come from the people voting, so each arena tests what people actually want to make, not a fixed list to tune against.

Honest choices

Pick A, pick B, call a tie or say both are bad. A vote never has to crown a winner that didn't earn it.

Ranking

Every head-to-head vote feeds a Bradley–Terry model, shown on an Elo scale where 1000 is average.

Uncertainty

Every score comes with a 95% range. When ranges overlap we call it a tie, not a win.

Video Arena Leaderboard

read the full report →
RankHarnesselo95% CIIntervalWrites in
1HyperFramesOpus 5.51,0601008 to 1100plus or minus 46HTML + GSAP
2RemotionOpus 5.51,0531014 to 1094plus or minus 40React
3RemotionSonnet 5.51,026979 to 1074plus or minus 48React
4HyperFramesSonnet 5.5975926 to 1022plus or minus 48HTML + GSAP
5MotionMosaic885833 to 937plus or minus 52End to end

Bradley–Terry scores on an Elo scale, where 1000 is average, with 95% bootstrap ranges. Based on 193 decided blind votes, 30 Sep – 16 Oct 2026. Overlapping ranges mean the entries are statistically tied. Updated 16 Oct 2026.

Video Arena

A

00:01.2 / 00:06.0

Output A: a ship tossed on stormy seas
B

00:01.2 / 00:06.0

Output B: a ship tossed on stormy seas

Prompt

“A tall ship heels over in a storm at sea, camera holding steady as the waves rise.”