Google's Veo, Alibaba's Wan and Pruna AI Are Now Fighting Over Cents Per Second

Every model launch this month came with a demo reel and a price sheet, and the price sheet is the one worth reading twice. Judged on render quality, nothing that shipped in September looks like a leap. Judged on what it costs to make a second of finished video, almost everything moved.

The obvious counterargument is that price has always been the least interesting part of an AI video launch. Creators pick tools by what survives a scroll on their feed, not by a rate card, and a cheap model that mangles hands is worth nothing. That view deserves a fair hearing before it gets argued down.

But the numbers on live pricing pages say the argument has quietly moved. What a creator can build this weekend is decided less by whose demo reel is best and more by whose per-second rate a budget can survive.

The Case for Ignoring Price and Judging Models Purely on Quality

The steelman is straightforward: a video model is a creative tool, chosen for what it produces, not what it bills. A five-second clip that needs three regenerations to fix melting fingers is not cheap at any price, and a model that nails a brief on the first try is not expensive at any price either. Under that logic, the interesting story each month is still the leaderboard, not the invoice.

Undominated.ai's independent video leaderboard backs that up: it ranks models by Arena-style votes first and price second, and Pruna AI's own p-video sits at roughly rank 27 with an Elo-style score of 1207.2 - well below ByteDance's Seedance 2.5 at 1482.0, priced at $0.30 per second on Runway's hosting. A model that is cheap per second but weak on the leaderboard has not made anything more affordable; it has just moved the retry cost somewhere less visible.

Google's own pricing structure implicitly makes the same case. The Gemini API's Veo 3.1 pricing page charges $0.60 per second for Standard 4K and $0.05 per second for Lite 720p - a twelvefold spread inside one model family, because Google is pricing capability tiers, not seconds in the abstract. If quality didn't matter to the price, that spread would not exist.

Promotional pricing complicates any headline number, too. Alibaba Cloud's own Wan 3.0 launch page says the flagship's "Standard" tier is 30% off through September 24, 2026. Pruna AI is running a parallel 50%-off window on P-Video-2-Pro through the same date, per Runtime Wire's report on the launch. A rate that expires in days is this week's price wearing a permanent-looking label, not this month's.

That is a genuinely strong case for waiting on the demo reel over the invoice.

Why the Per-Second Price Is Still the Real Story

The turn is this: the discounts themselves are the story, because they reveal how thin the margin between models has become.

Look at what Alibaba's own catalog shows on one page, the Model Studio pricing document. Wan 2.7's text-to-video model lists at $0.10 per second at 720p and $0.15 at 1080p, no discount attached. Wan 3.0's list price is actually higher before any promotion - $0.10 at 720p, $0.20 at 1080p - because it renders natively up to 30 seconds with omni-reference inputs Wan 2.7 lacks. Only the 30%-off window drags Wan 3.0's 1080p rate down to about $0.14, a hair under Wan 2.7's $0.15. When that discount lapses September 24, the newer, more capable model reverts to costing more per second than the one it replaced. The "collapse" in this pair is a launch promo, not durable repricing - and that distinction tells a creator exactly what to budget for next week.

Meanwhile the floor under the whole market keeps dropping regardless of what any one flagship charges. Pruna AI's own launch thread on X prices P-Video-2-Pro, an API wrapper around MiniMax's open-weight H3 model, from $0.02 per second in 480p Speed mode to $0.075 in 768p Quality mode - full price, before the current 50% cut. Pruna's own developer portal publishes the same ladder with the same terms - $0.02 for 480p Speed, $0.04 for 480p Quality, $0.035 for 768p Speed and $0.075 for 768p Quality, 50% off until September 24, 2026 - and Replicate's own post advertises the discounted floor at $0.01 per second for its first week.

The arithmetic makes the stakes concrete. Rendering 1,000 five-second clips on Veo 3.1 Standard at 1080p, at Google's listed $0.40 per second, costs 1,000 × 5 × $0.40 = $2,000. The same 1,000 clips on Pruna's cheapest full-price Speed tier, at $0.02 per second, cost $100 - a twentyfold gap between two API routes open today. That gap exists because an infrastructure shop hosting somebody else's open-weight model can undercut a lab that trained and owns its own, and creators choosing an API this month are choosing which side of it they stand on.

None of this makes the cheap route right; a $100 batch of clips nobody watches is still a wasted $100. But the question a creator should ask is no longer only "which model looks best" - it is "what does a second of usable video cost from this vendor, this week, before the discount ends." Platforms have noticed, too: resellers like Wan 3.0 Studio now sell one shared credit pool across Wan 3.0, Wan 2.7 and Seedance rather than a model-by-model rate card, which only works if the underlying per-second costs are close enough to blend.

Watch three things over the next month: whether Wan 3.0's price reverts upward on September 24 as its own catalog page says it will, whether Pruna publishes a real list price once its 50%-off window closes, and whether any outlet runs a real quality-per-dollar test rather than a quality test and a price sheet side by side. Credit-pool pricing bets creators won't notice which model actually rendered their clip. A vendor confident in its per-second price says so on the catalog page; one that isn't buries it in a promo.

Check today's live rate on the vendor's own pricing page before budgeting a shoot - not last week's screenshot, and not this article's numbers next month.