revid.ai logo

Case study · September 2026

We timed 713,649 AI video clips

How long Seedance, MiniMax Hailuo, Veo 3.1 and Sora 2 really take, how often they refuse, and why we switched models. From Revid production traffic, Jun 29 to Sep 26, 2026.

Save the headline stat as an image

23s

median clip on MiniMax H3 Max Turbo

the fastest model we run

12×

faster pro clips after leaving Seedance 2.0 Mini

35%

of Seedance 2.5 calls failed

mostly a real-person filter

Made with Revid · 34s · 1 prompt · 173 credits

We gave Revid the link to this page. It made the video.

Revid read the study, picked the takeaways, wrote the script, voiced it and animated every scene. Remake it with the same settings, or point the brief at your own article.
The prompt
https://www.revid.ai/case-studies/ai-video-model-index Create a less-than-one-minute video that shows the best takeaways from this article. It will be displayed on top of the article as a summary of the article, start with a great hook. Use the exact numbers from the article.

Motion graphics · Ultra · Minimal & type style · 16:9 · Alex voice · No captions · Startup Energy music

Remake this video Use it on your article

Speed

From 23 seconds to 7 minutes per clip

Click to finished clip, end to end. Grey rows are the models we replaced. Hover a row for details.

Swipe sideways to see the whole chart

10s30s1m2m5m10m20m40mBase, nowMiniMax H3 Max Turbo23sBase, beforeSeedance 1.0 and 1.51m 9sPro, nowMiniMax H3 Max31sPro, beforeSeedance 2.0 Mini6m 14sVeo 3.1Google1m 31sSora 2OpenAI2m 27sUltra, beforeSeedance 2.0 Fast5m 3sUltra, nowSeedance 2.56m 40stime to generate one clip (log scale)
middle half of generationsmedianup to the 90th percentile
  • Starting from an image halves the wait: 18s instead of 39s on base.
  • Seedance 2.5 is the slow one: 27% of its clips take more than ten minutes.

How long will my clips take, and what will they cost?

Real medians from Revid production, end to end: prompt, optional source image, render and upscale.

Quality tier or model

Runs on MiniMax Hailuo H3 Max Turbo

Clip length

Start from

Typical wait per clip

18s

the fastest option here

Credits for the video

90

15 credits per 5 seconds × 30s

Clips done within a minute

85%

1 clip in 10 takes over

1m 33s

19.2k Base generations timed, Sep 4 to Sep 26. Image vs text split measured Sep 4 to 26.

The switch

Seedance made 99% of our clips in the spring. MiniMax Hailuo made 86% in September.

Share of clips by model family. Users pick a tier, Revid picks the model, so this is our routing, not a popularity vote.

Swipe sideways to see the whole chart

0%25%50%75%100%JanFebMarAprMayJunJulAugSep*
  • Pro moved on August 27, base on September 3: 12× and 3× faster at the same price.
  • Seedance 2.5 still runs the ultra tier, for quality and clips up to 30 seconds.

Failures

The newest models refuse the most

Share of calls to each model that returned an error, before Revid retries on another host.

  1. Seedance 1.0 Pro Fast

    0.18%

  2. Seedance 1.5 Pro

    0.73%

  3. MiniMax H3 Max Turbo

    1.3%

  4. MiniMax H3 Max

    1.6%

  5. Seedance 2.0 Mini

    4.8%

  6. PixVerse V6

    6.3%

  7. Grok Imagine Video 1.5

    13.5%

  8. MiniMax H3

    14.9%

  9. Kling

    26.3%

  10. Seedance 2.0 Fast

    32.2%

  11. Seedance 2.5

    35.1%

Why calls fail

Swipe sideways to see the whole chart

0%25%50%75%100%SeedanceH3 MaxHailuo H3PixVerseGrokKling
  • 55% of Seedance failures are the same message: “input image may contain a real person”. Use an illustrated or 3D character instead of a photo.
  • Same model, different host: Seedance 2.5 failed 51% of calls on the maker's API and 5% elsewhere.

Format

More people make vertical clips. Landscape makers make more of them.

Share of Seedance clips

  1. 16:9

    55%

  2. 9:16

    39%

  3. 1:1

    4%

Creators using each ratio

  1. 16:9

    3,186

  2. 9:16

    6,466

  3. 1:1

    991

  • A landscape creator generates 2.8× as many clips as a vertical one (90 vs 32).

All models

Side by side

Sort by any column.

ModelMain reason calls fail

Seedance 1.0 Pro Fast

ByteDance

37.4%3,9040.18%1m 9s*15Rejected, reason not returned (56%)

Seedance 1.5 Pro

ByteDance

23.1%3,9730.73%1m 9s*15Content refusal (50%)

MiniMax Hailuo H3 Max Turbo

MiniMax

15%8,5021.3%23s15Rejected, reason not returned (90%)

Seedance 2.0 Mini

ByteDance

11.4%4,3634.8%6m 14s50Content refusal (83%)

MiniMax Hailuo H3 Max

MiniMax

4.7%4,1401.6%31s50Rejected, reason not returned (90%)

PixVerse V6

PixVerse

3.8%1,3476.3%–50Rejected, reason not returned (96%)

Seedance 2.5

ByteDance

1.6%44135.1%6m 40s100Content refusal (98%)

MiniMax Hailuo H3

MiniMax

1.1%79714.9%–100Rejected, reason not returned (39%)

Seedance 2.0 Fast

ByteDance

0.83%25032.2%5m 3s100Content refusal (89%)

Grok Imagine Video 1.5

xAI

0.64%75813.5%–15Content refusal (75%)

Kling (O3 reference, 3.0, 2.1)

Kuaishou

0.17%31926.3%––Rejected, reason not returned (98%)

Google Veo 3.1

Google

–––1m 31s100–

OpenAI Sora 2

OpenAI

–––2m 27s130–

Failed calls count every provider call that returned an error, before Revid retries on another host, so most never reach the user. * Seedance 1.0 Pro Fast and 1.5 Pro shared the base tier and are timed together. Veo 3.1 and Sora 2 are timed but not counted in shares.

Made with Revid

Pick a tier. Revid picks the fastest model that can do it.

Revid routes every clip to the model that fits and retries on another host when one refuses.

Make a video freeAI video generator

Free to start · no credit card

How we measured this

  • Source: Revid's production analytics, Jun 29 to Sep 26, 2026 (UTC). A clip is one successful call to a video model. 713,649 clips for 21,636 creators; internal and API accounts are included. The monthly share chart uses the same events from January 2026.
  • Share reflects Revid's routing: users choose a tier or a named model, Revid maps each tier to a model per route (image to video, first and last frame, native audio, reference) and falls back to other hosts on failure. Routing changed several times inside the window, so per-model totals mix periods.
  • Failed calls are counted per provider call, not per user request. Fallbacks mean a creator can see one clip after several failed calls, and a failed call is often rescued by the next host. 2.9% of all calls failed. Reasons come from rules on the logged error message; a bare "unprocessable" rejection is reported on its own rather than guessed, though a two-hour sample on September 7 found all 20 rejected MiniMax Hailuo H3 Max calls were content-policy refusals.
  • Speed is end-to-end wall-clock time of single-clip generations (the /create media step and editor regenerations), logged per tier and attributed to the model the tier ran on in each date window. Percentiles are interpolated from 5 to 40 second histogram bins; the median agrees with a direct median within 0.4 seconds. The image vs text split was measured September 4 to 26. Speed moves day to day: the base median sat at 15 to 19 seconds from September 7 to 23, then rose to 32 to 38 seconds on September 24 to 26.
  • Veo 3.1 and Sora 2 appear only in speed (not in share or failures), with small samples (295 and 129 clips). Aspect ratio is only known when the call carried one; MiniMax Hailuo H3 Max image-to-video follows the source frame and logs none. Credits are the price list on September 28, 2026; pro and ultra prices were cut on August 12, inside the window.

Data pulled September 28, 2026. The index grows every day; numbers are refreshed on a schedule and the date above always tells you which snapshot you are reading.

Cite this study

Revid (2026). AI Video Model Index 2026: How Fast and Reliable Seedance, MiniMax Hailuo and Kling Really Are (713k Generations). https://www.revid.ai/case-studies/ai-video-model-index

More from the Revid index

🎵Case study

11

AI music video generators scored on 10 checks

Best AI music video generators

We scored 11 AI music video generators on 10 things a musician needs, from full-song output to lip sync and publishing, with a source for every claim. Plus what ChatGPT and Google AI Overviews recommend, and why the most-named tool cannot take a song.

Read the study
🎞️Case study

3.5×

median likes when the post names Claude

Claude motion graphics, measured

Frame-by-frame analysis of the 63 Claude Opus 5.5 motion graphics videos that passed 100 likes on X in launch week: cuts, color, motion, sound and openings, against 13M views and 74k likes. What actually separated the hits.

Read the study
📼Case study

1.55×

median views for AI-made nostalgia vs real footage

The 80s AI video trend

Google searches for 80s AI video prompts went from almost nothing to 1,945 impressions in a day in September 2026. We read that spike against 1,676 80s TikToks: formats, views, likes, and how AI-made nostalgia performs against real footage.

Read the study
💘Case study

7×

more views for couples content than crypto

Love beats money

We measured the average views of 46 hashtag themes across 2.4 million TikToks. Couples content averages 7x the views of crypto. Here is the full ranking of what TikTok rewards, from romance to real estate.

Read the study
🫣Case study

2.2×

more views for handheld footage than selfie videos

Your face costs views

Selfie-style videos average half the views of handheld footage, faceless videos beat person-focused ones, and animals beat everyone. The data on whether you need to show your face on TikTok.

Read the study
🤖Case study

2×

average views for AI-generated TikToks vs real footage

AI slop vs real footage

In our index, AI-generated TikToks averaged twice the views of real footage since May 2025. On X, Instagram and Threads, AI visuals are 5x more common in the lowest engagement tier than in the viral tier. Both are true.

Read the study
🩺Case study

3,000

viral health TikToks audited from their transcripts

TikTok health claims

An audit of the 3,000 most-viewed English health TikToks in our index: how many make a health claim, how many back it with anything, how many sell a product, and which claims went most viral.

Read the study
FAQs

Frequently Asked Questions

Which AI video model is the fastest?
In Revid production from Jun 29 to Sep 26, 2026, MiniMax Hailuo H3 Max Turbo was the fastest: a median of 23s from click to finished clip, with 59% done in under 30 seconds. MiniMax Hailuo H3 Max took 31s, Veo 3.1 1m 31s, Sora 2 2m 27s and Seedance 2.5 6m 40s. Times are end to end and include prompt writing, any source image, rendering and upscaling.
How long does Seedance 2.5 take to generate a video?
A median of 6m 40s per clip across 2,058 timed generations, with the middle half between 4m 40s and 10m 27s. One clip in ten took longer than 16m 51s, and 27% ran past ten minutes.
Why does Seedance say the input image may contain a real person?
Seedance runs a privacy filter that refuses source images it thinks show a real person. It was the single biggest reason Seedance calls failed on Revid: 8,218 of 14,979 failures (55%), almost all on the model maker's own API. On that API, 51% of Seedance 2.5 calls failed, against 5% through a third-party host. Starting from an illustrated or 3D character instead of a photo of a real person usually gets past it.
Is MiniMax Hailuo better than Seedance?
On speed and reliability in this data, yes for the everyday tiers: MiniMax Hailuo H3 Max cut the median pro clip from 6m 14s on Seedance 2.0 Mini to 31s, and failed 1.6% of calls. Seedance 2.5 remains the top-quality option and runs Revid's ultra tier, at the cost of long waits and frequent refusals of images with real people.
How much does AI video generation cost on Revid?
Revid bills credits per 5 seconds of generated video: 15 on the base tier (MiniMax Hailuo H3 Max Turbo), 50 on pro (MiniMax Hailuo H3 Max), 100 on ultra (Seedance 2.5), 100 for Veo 3.1 and 130 for Sora 2. Seedance 2.5 at 1080p costs 250 and 4K costs 500. Prices as of September 28, 2026.
How was this measured?
From Revid's production analytics: every successful or failed call to an AI video model between Jun 29 and Sep 26, 2026 (713,649 successful clips for 21,636 creators), and the end-to-end generation time of every single-clip generation. Share reflects Revid's own routing, which changed during the window, not a neutral market share.

Still have questions?

Contact us

Find detailed answers to 100+ questions about features, tools, and workflows

or check our markdown version optimized for LLMs