Veo 3.1 still wins always-on audio
Eleventh on some Elo boards and still the pick when audio cannot be optional. Official 4K tier. Timestamp-block prompting is Google's multi-shot method.
Veo 3.1 will lose an arena week. It has already lost several. That is not a reason to abandon it. The job it owns did not move: audio that is on unless you turn it off, dialogue that is part of the generate, and a 4K file when the delivery sheet asks for one.
Elo is a vote on a short clip. Always-on audio is a contract. Veo 3.1 is still the model you open when the brief cannot survive a silent plate.
What "always-on" actually means here
It does not mean the model refuses to render quiet. The catalog exposes an audio on/off matrix: 20 credits a second with audio off, 40 with audio on, durations 4 / 6 / 8 seconds, qualities 720p / 1080p / 4K. Always-on means the default professional path is the audio-on row, and turning it off is a choice you make for a silent plate — not the way the model "usually" arrives.
Google's Veo 3 family generates 48 kHz native audio in the same pass as the picture. Speech, room, and effects are not a second product you hope to lip-sync later. That is the job. A model that can emit a soundtrack on a pretty 8-second clip is a different product. The arena currently rewards that second product. Your media buyer does not.
Three tells you are still in Veo's lane:
- The mouth is the shot. Founder, presenter, UGC talent, a line of copy that has to be said, not captioned.
- Silence is a defect. If a muted export would fail the brief, you need native audio, not a bed you will add in the edit.
- The master is 4K. Veo lists 4K as an output quality. Most of the models beating it on Elo do not.
If none of those are true, Veo is an expensive way to make a landscape. Pick the row that matches. The audio filter across the catalog lives at best model with audio; Veo is a named member of that set, not the only one, and it is the one you want when the line is the brief.
Timestamp blocks are the multi-shot method
Google's documented way to get more than one beat out of a Veo generate is not a prose paragraph. It is timed blocks: 0–2s, 2–5s, 5–8s, each with a shot, an action, and — if you need it — a line.
That is a directing tool inside an 8-second ceiling, not a film. You are not getting Seedance's 30-second single pass and you are not getting Kling Omni's storyboarded cuts as a different architecture. You are getting one short generate whose internal time is labelled. Write the labels. The longer craft note is timestamp blocks in Veo 3.1 prompts; the decision rule here is simpler: if you need two shots and a line inside eight seconds, block the clock. If you need a new room, generate a new clip.
A paragraph that says "then she turns and then she speaks and then the camera rises" is how Veo compresses beats and paraphrases dialogue. One action per block, the spoken line in quotes, "no music" when you do not want a score. That is the whole prompt discipline for this model.
What Veo is not
It is not a 30-second identity lock. Durations stop at 8 seconds. Continuation is a different call (extend when the last frame is good; regenerate when the idea is wrong). Do not stretch a talking shot by asking Veo to "keep going" in prose.
It is not the I2V pack-shot model. Seedance 2.5 owns still-to-motion lock and the fat reference budget. Ingredients-style references on Veo exist and they are a different, smaller contract. If the product has to match the still, start on Seedance.
It is not a volume hook engine. Eight seconds of 4K audio-on Veo is a hero cost. Thirty variants of a silent opening belong on a cheaper, faster row.
It is not the arena leader. Treat every "Veo dropped to eleventh" screenshot as weather. Weather is not a re-stack.
Keep it on the roster on purpose
The failure mode after a lost arena week is aesthetic panic: the team regenerates last month's talking ads on whichever model is currently pretty, then spends the edit putting a voice back on a mouth that was never built around the line.
Hold the lane instead:
- Talking shots, localisation of a spoken line, any brief where the soundtrack is the product — Veo, audio on.
- Silent plates you will score later — Veo audio off, or a model that does not pretend to speak.
- Pretty 8-second native-audio B-roll — Omni Flash is allowed to win that. That is not a talking ad.
You already have Gemini Omni Flash leads the arena, not the dialogue job if you need the board explained. This page is the operating instruction: do not abandon Veo because it lost a week. The job it owns is still the job.
FAQ
Is Veo 3.1 still worth it if it ranks below Omni Flash and H3?
Yes, for the dialogue and 4K jobs. Rank is an 8-second preference vote. Veo's catalog contract is native audio as a billed toggle, a 4K quality, and a duration set built like a talking shot. Those did not expire when the board moved.
Does Veo always generate audio?
No. Audio off is a real setting at a lower credit rate. Use it when you want a silent plate. "Always-on" is the professional default for spoken work, not a missing mute button.
Can I prompt a whole scene in one paragraph?
You can. You should not. Google's own multi-shot method is timestamped blocks inside the 8-second window. One action and one line per block. A paragraph of "and then" is how beats disappear.
When should I leave Veo for Seedance or Kling?
Leave for Seedance when identity has to lock off a still or a reference stack, especially toward 30 seconds. Leave for Kling when the value is motion across cuts, not a mouth. Stay on Veo when the line has to be said.
Arena weeks are cheap to lose. A talking ad generated on a model that does not own speech is expensive to redo. Keep Veo on the dialogue row until the contract changes, not until the leaderboard does.