AI Product Video at Catalog Scale

AI product video crossed a pricing threshold in the first week of August 2026: the models that generate short product clips are now quoted per second, on published rate cards, in the same units you already use to budget API spend. That reframes catalog-scale video from a production question into an arithmetic one. A hundred SKUs at six seconds a clip is a two-figure line item — which means the hard part has moved somewhere else entirely.

Two of those rate cards landed inside a fortnight. Black Forest Labs took FLUX 3 Video to general availability on August 4, 2026 with vendor list pricing from $0.17 per second, a figure its OpenRouter listing matched the same day. MiniMax shipped H3 on July 31 with vendor pay-as-you-go pricing at $0.13 per second for 2K and $0.09 for 768P. Kling publishes a full per-second rate card too, but in yuan only. And ByteDance launched Seedance 2.5 on July 31 without publishing a per- second price at all.

This playbook does the thing none of the vendor pages do: it normalises those rates into one catalog-scale cost table, maps the pipeline from an existing product-detail-page image to a published ad, and gives equal billing to the failure modes — hands, on-screen text, and physics — that still make a human QA gate mandatory. If you have already worked through AI product photography for still images, this is the moving-picture sequel, and the economics are not the same shape.

For as long as product video has existed, its unit of cost has been a day — a studio day, a crew day, an edit day. That unit is indivisible and it does not scale down, which is why catalog-scale video has historically been reserved for hero SKUs and seasonal campaigns. Everything else got a still image and a bullet list.

Three of the four models below broke that unit into seconds — the fourth is the counterexample this post keeps coming back to. Not because the output is equivalent to a studio shoot — it is not, and section 05 is entirely about where it falls short — but because a divisible unit lets you ask a different question. Instead of “which twelve SKUs deserve video,” the question becomes “what does it cost to give all 400 of them a six-second clip, and what does the review process cost on top.”

Worth separating two events that coverage keeps merging: the FLUX 3 announcement in late July was early access to the base FLUX 3 image and multimodal family, and it is a different thing from the August 4 video GA. The 20-second clips, the native audio, and Draft Mode all belong to the video release. Our FLUX 3 Video GA guide covers that launch in depth, and the MiniMax H3 launch and Seedance 2.5 one-take launch posts do the same for the other two.

Every vendor quotes its own number, in its own currency, on its own surface. Nobody publishes the comparison, so we built it. The table below takes a fixed workload — 100 SKUs, one six-second clip each — and applies each published rate to it. Every cell is arithmetic on a rate we could read on a primary page at the time of writing: price per second × 6 for the clip, × 100 for the run. No efficiency claims, no discounts, no volume tiers folded in.

Read the surface label on every row, because they are not interchangeable. A vendor list price is what the model owner charges directly; an aggregator listing is what a routing layer charges to reach the same model. On FLUX 3 Video the two agree on the base rate quoted here — but that is something to check per model and per tier rather than assume, and either surface can change without an announcement.

Two things jump out of that grid. First, the USD spread across the comparable tiers is narrow: $54.00 to $102.00 for the same 100-SKU, six-second workload. At that spread, model choice should be driven by output quality and reference handling, not by price — a $48 difference across an entire catalog run is noise next to one round of reshoots. Second, the 4K silent tier on Kling is roughly five times the 720P silent tier per second, which is the only place in this table where a resolution decision materially moves the budget.

One more surface note, because it changes how you shop. OpenRouter is not a universal price-comparison layer for video the way it is for text models — at the time of writing its listing carried FLUX 3 Video but not Kling 3.0, MiniMax H3 or Seedance 2.5, each of which is API-direct or vendor-platform only. If you are used to comparing language models on one aggregator page, expect to visit three or four vendor consoles instead. The broader dynamics of that market are covered in our piece on the AI video price war.

The pipeline that actually works at catalog scale does not start from a text prompt. It starts from the photography you already have on the product detail page, because that is the only asset that guarantees the generated clip shows your product rather than a plausible cousin of it. Every model in this comparison supports image-to-video or reference-to-video for exactly this reason.

Below is the stage map, with the model feature that serves each stage and the point where that stage typically breaks. It is not a vendor diagram; it is the sequence a team ends up running once the first batch comes back with problems.

Consistency is the whole ballgame in ecommerce video. A model that produces a beautiful clip of something that is 90% your product has produced nothing you can publish. The lever the three priced models give you is the reference system: how many anchoring inputs you can supply, and of what kind. They differ enough that this alone can decide the model.

The practical read: if your catalog has strong existing photography and each SKU needs one clean clip, keyframe-driven image-to-video is enough and FLUX 3 Video’s Draft Mode is the cost lever. If your SKUs are visually similar to each other — variants of the same silhouette in different finishes — you want the widest reference set you can get, because that is what stops the model averaging across your own range. And if the plan is one master creative with the product swapped per variant, that is a different technique entirely, covered in our guide to swapping product creative for local ad variants.

Here is the honest framing, and it needs stating precisely because it is easy to overclaim in both directions. There is no published benchmark for product-video accuracy. None of the vendors scores hands, small on-screen text, or physical plausibility; none claims those problems are solved. What exists is a pattern visible across the vendors’ own launch materials and early independent coverage — the same categories of artefact keep appearing, across models.

Treat the list below as a QA checklist, not as a measured error rate. The distinction matters when you are setting expectations with a merchandising team: “these are the things to look for” is defensible; “model X fails 12% of the time on reflective surfaces” is a number nobody has published.

The three recurring categories

• Hands. Any clip where a person holds, opens, or demonstrates the product puts hands in frame at close range — the single most scrutinised region in generated video, and the one a customer will notice first.
• On-screen text. Labels, ingredient panels, model numbers, and brand marks on packaging. This is the failure mode with the highest commercial consequence, because a garbled ingredient list on a consumable is not an aesthetic problem.
• Physics. Liquid pours, fabric drape, and reflective or transparent surfaces. Glass, chrome, foil packaging and anything with a specular highlight are where implausible motion is most visible.

Section 02 priced generation. This section prices the part that decides whether the project is worth running. The arithmetic below is planning arithmetic on stated assumptions — three minutes of human review per clip, an illustrative $40/hour blended internal rate, and a quarter of clips failing the gate and needing one regeneration. Those assumptions are yours to argue with; substitute your own numbers and the shape of the answer holds.

That ratio is the finding, and it inverts the intuition most teams bring to the decision. The instinct is to shop hard on per-second price; the arithmetic says the per-second price is roughly a third of the true cost and the spread between the cheapest and most expensive comparable model is $48 across an entire 100-SKU run. The variable that actually moves your budget is the failure rate, and the variable that moves the failure rate is the quality and consistency of your source images — stage 01, the one that costs nothing but attention.

Which suggests a different procurement question. Not “which model is cheapest per second,” but “which model’s reference system most reduces the number of clips my reviewers reject.” A model that costs $0.08 more per second and halves your rejection rate is straightforwardly the cheaper option on this arithmetic. That is not a question any rate card answers, and it is the one worth running a paid pilot on — twenty SKUs across two models, one reviewer, same checklist. It is the shape of evaluation our ecommerce engagements start with.

A generated clip is not a deliverable until it fits the surface it runs on, and the surfaces disagree with each other — and, in one case, with the models. Shopify’s own product-video guidance puts the sweet spot at 30 to 60 seconds, with 15 seconds to 2 minutes as the usable range. FLUX 3 Video caps a single pass at 20 seconds and MiniMax H3 at 15. That gap is not a flaw in either; it just means a 30-second product-page video is a stitched sequence or a multi-round extension, not one generation.

One clip therefore serves several surfaces if — and only if — you re-crop it in post rather than re-render it per placement. Regenerating per aspect ratio multiplies both the generation bill and, more painfully, the QA load, because every re-render is a new clip that has to clear the gate. For the short-form end of the funnel, our guide to short-form product clips on YouTube Shorts covers the platform-specific side, and if 4K masters are the goal, Kling 3’s 4K rendering is the tier to price against. On the buying side, the placement mechanics sit with paid media, not with the generation stack.

The demand-side evidence for product video is widely quoted and weakly sourced, and it is worth being precise about what it is. A 2026 Wyzowl marketing survey reports that 85% of people say they have been convinced to buy a product or service by watching a video, with 63% preferring a short video to learn about a product against 12% for text articles. A Shopify guide puts the equivalent purchase-influence figure at 89%, and reports that 96% have watched an explainer video to learn about a product or service.

Both are marketing content produced by a video-marketing vendor and an ecommerce platform respectively — not independently audited research, and not academic work. Nor can we show that the two are independent of each other: the figures land close together, but a platform guide quoting a purchase-influence number may be restating the same vendor survey, which would make the agreement one origin counted twice rather than two readings that match. Treat it as a single directional signal. What it is not useful for is a conversion forecast for your catalog. Use it to justify a pilot, never to size a business case.

The defensible business case is the one you can compute from the numbers in section 06. A 100-SKU run costs on the order of $377.50 all-in on the stated assumptions. Against that, the only question that matters is whether the clips move the metric on the pages they sit on — which is a measurable thing, on your own traffic, with a holdout set of SKUs that stay video-free. That test costs the same $377.50 and produces a number that belongs to you rather than to a vendor’s survey panel. Wire it into the PDP itself as a controlled experiment and the result generalises to the rest of the catalog.

Where this goes next

Two pressures are visible in the current field and both point the same way. First, per-second rates across the comparable tiers have converged into a narrow band, which is what happens when a capability commoditises; the differentiation is migrating to reference handling, editing of existing clips, and cost-control mechanics like draft renders. Second, the single-pass ceiling — 15 to 30 seconds today — sits just below the length the distribution surfaces actually want, which is an obvious gap for the next release cycle to close.

If both hold, the bottleneck in catalog-scale product video will not be model access or price by the end of 2026. It will be review throughput: the human hours between a rendered clip and a published one. Teams that invest now in a tight, checklist-driven QA gate and clean, normalised source photography will be able to absorb the next price drop as pure volume. Teams that treat generation cost as the project will find that the cheaper the seconds get, the more the review queue becomes the entire problem.