H3 vs Omni Flash is an arena tie

Gemini Omni Flash at 1245 Elo and MiniMax-H3 at 1242 on Artificial Analysis text-to-video with audio is a tie, not a verdict. invideo’s August 2026 roundup (Best AI Video Model in August 2026, updated 7 August 2026) cites that blind-arena snapshot: Omni Flash leads T2V+audio at 1245, three points over MiniMax-H3 at 1242, with Seedance 2.0 third at 1225. Three points is inside a beauty contest’s noise. The job split is not Elo. On Versely you pick Gemini Omni Video / Gemini Omni Flash Edit or Minimax H3 Text to Video because the card matches the shot, not because a leaderboard blinked.

This page is the arena correction. It is not “Omni Flash leads, therefore skip H3,” and it is not a dialogue bake-off with Veo.

What the August board actually scored

A T2V-with-audio arena asks a voter to watch two short clips and pick a preference. invideo’s table for August 2026, attributed to Artificial Analysis:

Rank	Model	Elo (Aug 2026)
1	Gemini Omni Flash	1245
2	MiniMax-H3	1242
3	Seedance 2.0	1225

invideo’s own headline after the table is the useful one: standings move weekly, so the durable ranking is by job. Omni Flash’s documented weak spot in that roundup, by Google’s own admission as quoted there, is short speech segments — score dialogue elsewhere. H3’s pitch in the same piece is high resolution with references: 2K native, up to 9 image / 3 video / 3 audio references on MiniMax’s own stack. Those are job sentences. Elo is a vote on an 8-second-ish file.

What the exam does not ask:

• Whether the mouth holds a specific line for an ad.
• Whether you can turn audio off.
• Whether 2K is on the card you will actually confirm.
• Whether a later edit pass can change the jacket without resampling.
• Whether 15 seconds is available, or only 10.

Three points at the top of a blind board means “voters did not agree these are different species.” It does not mean “always Omni.” It does not mean “always H3.” If your workflow is “whoever is #1 this morning,” you will churn rows every time the board refreshes and never learn why a clip failed.

Other snapshots the same month show the same tie with different integers: coverage of H3’s open weights in early August put T2V+audio in a dead heat (Omni slightly ahead, intervals overlapping). Do not fight over 1245 versus 1243. Cite invideo’s August roundup when you need a dated public number, then ignore the third digit.

What the two Versely cards actually are

Omni family, two rows. Gemini Omni Video: T2V + I2V + R2V, 4s / 6s / 8s / 10s, 720p / 1080p / 4K, from 29 credits, quota images + videos*2 + character_ids <= 7. Gemini Omni Flash Edit: video-to-video retouch, from 50 credits, audio off on that card. Omni 1.1 Flash (gemini-omni-1.1-flash, Google 27 August 2026) is a Google-surface update — 10s extend memory, 40s cap, 360p drafts, first-last, 3s video ref — not a third Versely slug. Arena “Omni Flash” is the family voters saw. Your confirm button is still one of the two published rows, or Google’s 1.1 door.

H3, one published generate. Minimax H3 Text to Video: text-to-video, 5s through 15s every second, 2K max output, from 104 credits, no image required. Catalog audio field is unset. Aspects include 21:9 through 9:16. Qualities include 768P and 2K. MiniMax’s broader H3 stack (9 stills, 3 video refs, 3 audio refs) is vendor capability. This Versely card is the text-to-video generate. Do not brief nine reference images on a row that does not ask for them.

Credit comparison is not Elo. Omni Video starts at 29 credits for up to 10 seconds and a 4K listing. H3 starts at 104 credits for up to 15 seconds and a 2K listing. Different length, different resolution ceiling, different price. A 15-second 2K plate is an H3-shaped job even if Omni is three Elo ahead on a short-clip board. A 6-second Omni generate you plan to retouch in Edit is an Omni-shaped job even if H3 won a different snapshot.

Seedance 2.0 sitting third at 1225 is the reminder that “below H3 on T2V+audio” can still be #1 on I2V in the same invideo roundup (Seedance 1196 on I2V+audio). Boards are per input. Do not global-rank a family from one column.

Split the job, ignore the third digit

Pick Omni Video when the shot is a short multimodal generate: prompt plus optional stills or clips under the quota, 4–10 seconds, and you want the Omni look you can later retouch on Flash Edit. Pick Omni when you already live in Google’s stack and 1.1’s draft/extend/first-last controls are in play on Google’s door. Pick Omni when a card that starts at 29 credits and caps at 10s matches the brief.

Pick Omni Flash Edit when the take exists and one thing is wrong. Elo does not retouch. Edit does.

Pick H3 when the shot is text-only, you want 2K, and 5–15 seconds is the length. A 12- or 15-second hold is inside H3’s menu and outside Omni Video’s 10s cap. Widescreen 21:9 is on H3’s aspect list. If the brief is “2K from a sentence, no still,” that sentence is the H3 card.

Pick neither when the job is always-on dialogue at Veo’s cinematic spec: Veo 3.1, 4 / 6 / 8s, 4K, from 64 credits, native audio. invideo’s roundup already sends speech elsewhere. Arena #1 is not a talking-head contract.

Pick neither when the job is a native 16–30 second single pass: Seedance 2.5, 720p, from 34 credits (was 67), audio on. H3 stops at 15s. Omni Video stops at 10s. 2.5 is the length row.

Pick neither when the job is a driving clip copied onto a character: Kling V3 motion control.

The AI video generator will show duration and resolution before you confirm. If the menu cannot hold the shot, Elo cannot either.

A practical bake-off, when you insist on one: same prompt, Omni Video at 8s, H3 at 8s (H3’s menu includes 8s). Watch on a phone, sound on. Keep the file you would ship. Do not keep the file whose family is currently #1 on a blog table. Run that test once per campaign, not once per hour.

After either file: captions. Arena clips are silent-autoplay in the feed even when the model wrote a track. Add captions to video. Label realistic synthetics for EU audiences: Article 50 disclosure.

Stop using the board as a router

Routers that say “#1 this week” produce three failures:

1. Length mismatch. You needed 15s, you bought 8s of the leader, you extended in an editor and called it the arena winner.
2. Resolution mismatch. You needed 2K from text, you bought Omni because of 1245, you upscaled and blamed H3 for a job you never ran.
3. Dialogue mismatch. You needed a line, you bought the T2V+audio leader, the mouth was pretty and wrong. Veo was the row. The board never asked about the line.

invideo’s August 2026 piece already ranked by job after printing the Elo table. Do that. Cite 1245 vs 1242 when someone asks “who won the arena.” Then open the card that matches duration, resolution, input type, and whether you will retouch.

Omni 1.1’s 27 August controls do not break the tie. They add extend memory, 360p drafts, first-last, and a 3s video ref on Google’s surface. H3’s 2K / 15s card does not get those controls by losing three Elo. Different doors, different extras.

If you want to run the two rows on one bill, sign up and confirm Omni Video or H3 from the duration menu — not from a screenshot of a leaderboard.

FAQ

Who is #1, Omni Flash or H3?

On Artificial Analysis T2V with audio as cited by invideo’s August 2026 roundup (updated 7 August 2026): Omni Flash 1245, MiniMax-H3 1242. That is a three-point gap on a blind short-clip vote. Treat it as a tie. Pick by job.

Does H3’s 104 credits mean it is “worse value” than Omni Video’s 29?

No. The cards are not the same product. H3: 5–15s, 2K, from 104 credits, text-to-video. Omni Video: 4–10s, up to 4K listed, from 29 credits, multimodal quota. Compare files that share length and input, not sticker prices across menus.

Should I use the arena leader for talking ads?

No. invideo’s own August writeup sends short speech elsewhere. On this catalog, Veo 3.1 is the cinematic dialogue-shaped generate (4 / 6 / 8s, 4K, from 64 credits, native audio). Elo is not a mouth contract.