Google prices a 40-second AI video sequence at four dollars

The ten-cent second is a division problem rather than a rate card entry. Google bills 5,792 output tokens for every second of 720p video, and video output is priced at $17.50 per million tokens ^[6], which comes to $0.1014 a second ^[2]. An unchanged headline rate is therefore an unchanged token rate ^[5]. For anyone building a plan, that yields two round numbers: a full 40-second sequence at 720p runs about $4.00 ^[3], and the same 40 seconds drafted at 360p runs about $1.33 ^[4].

List price is not the budget, and the one worked example in the record shows why. Higgsfield AI's Hell Grind, a 95-minute AI-generated science fiction feature that screened around Cannes in May 2026, was made in roughly two weeks for about $500,000, of which about $400,000 went to AI compute, and it still required a 15-person team ^[16]. Its first 25 minutes took more than 16,000 initial generations to arrive at 253 final shots ^[17], about 63 attempts per surviving shot ^[5]. Compute absorbed 80 percent of the budget ^[6], and $400,000 across 5,700 seconds of finished runtime is roughly $70 per delivered second ^[7], on different tooling in a different year, but the ratio is the lesson. The money is in the discard pile, which is what the draft tier is actually repricing.

A skeptic has a fair objection, and Ron Schmelzer, a Forbes contributor, put it plainly: Google's growing set of overlapping video products and entry points "could make the product story harder for users to follow" ^[18]. For a scheduling decision, though, the entry point matters less than whether the rate holds and whether the tool sits inside the workflow already in use, and WPP has integrated Omni Flash into WPP Open while Adobe plans to add it to Firefly ^[10]. Sprawl still carries a cost, and OpenAI has just shown its shape: the Sora product is no longer available, and the Sora 2 models and Videos API are deprecated with a shutdown date of September 24, 2026 ^[15]. Committing a production line to a video product is a bet on which entry point survives the vendor's own reorganisation.

The 40 seconds is cumulative, not continuous. A single generation still stops at 10 seconds ^[2], continuity comes from the model reading up to 10 seconds of prior footage instead of only the final second ^[3], and that state is carried forward by Google's stateful Interactions API, which passes the previous video and its references into the next request ^[8]. Meanwhile the model card still lists character consistency across edits and accurate text rendering as open problems ^[11]. The pilot worth running this quarter is therefore a test of the joins between increments and of any on-screen type, not of whether 40 seconds can be produced at all.

In board-deck terms, AI video became a schedulable line item at four dollars per 40 seconds. But the deck leaves out the delivery format: Google publishes no per-second rate for 360p, 1080p or 4K ^[7], so nobody can put a fixed cost on the file that actually ships. A team can commit to a drafting cadence this quarter and still be unable to sign a fixed-price deliverable next quarter, and that asymmetry is the trade this release offers.