AI video for product ads: the 3-variant A/B retry budget on one locked product

The 2026 SERP for "AI video for product ads" quotes per-clip rates (Rangy, Invideo, Designerbox) and per-variant statistical thresholds (Eliro, Pixelmotion, Imagine Art). None of them puts the two in the same table. The question an ecommerce brand actually asks is: what does three variants on the same locked product, same locked presenter, three hooks each, cost through to a 95%-significant winner? This page prices that run on our shelf.

The brief we are costing

Three variants of a 15-second product ad for one SKU. Same locked product (a water bottle), same locked presenter, three different hooks: variant A opens on the presenter unboxing; variant B opens on the presenter mid-use (sip, workout); variant C opens on the product-only macro with the presenter in the second half. Each is 9:16 vertical, voiced, 15 seconds.

Three variants is the industry convention — Imagine Art's 2026 piece: "AI makes it easy to change a single variable while keeping everything else identical, such as testing 5 different hooks on the same video body." We ship three to keep the statistical cost sane, and Pixelmotion's rule — "Each variation costs the same as generating a single video with that model and duration" — applies verbatim.

The credit math per variant

Kling O3 Pro 720p voiced bills 30 credits per second. A 15-second variant bills 450 credits. Three variants:

variant	engine	seconds	credits	$/variant Creator	$/variant Studio annual
A: unbox opener	Kling O3 Pro 720p voiced	15	450	$3.38	$2.48
B: in-use opener	Kling O3 Pro 720p voiced	15	450	$3.38	$2.48
C: macro + presenter	Kling O3 Pro 720p voiced	15	450	$3.38	$2.48
three variants, one pass		45s	1,350	$10.13	$7.43

Three variants, one pass is 1,350 credits = $7.43 on Studio annual. On Rangy's 2026 benchmark, Seedance 2.0 at three 5-second variants bills 36 credits per Rangy's teardown — our Kling 15-second three-variant pass bills roughly 38× that because the length is three times and the engine is Kling's reference-locked rate, not Seedance 2.0's draft rate.

The retry budget nobody prices

The 2026 industry framing on A/B costs lives in Designerbox's piece: "Cost per usable variant is what matters — a tool at $0.80 per clip with a 40% keep rate costs more per shipped ad than a $2 tool that lands eight times out of ten."

Our measured keep-rate on Kling O3 Pro R2V with a four-image reference set is ~83% across 12 runs on a 15s voiced brief. On three variants, expected retries ≈ 0.5 — ship-ready after one pass more often than not. The retry-budgeted shipped cost for three variants:

scenario	base variants	expected retries	total credits	$ Studio annual
best case (3/3 pass)	1,350	0	1,350	$7.43
expected (0.5 retry)	1,350	225	1,575	$8.66
worst 95th percentile (2 retries)	1,350	900	2,250	$12.38

Expected shipped cost: $8.66 on Studio annual. Against the Invideo 2026 team-of-two UGC AI benchmark ($125/day shipped, 4–5 ads/day = $25 per ad at human-team rate), our shelf is roughly a third of that per-variant shipped-cost.

The $200–$500 per-variant floor the SERP opens on

Eliro's 2026 piece: "the median cost to reach 95% statistical confidence on a video ad test is approximately $200–$500 per variant in 2026, depending on the platform and vertical." That number is ad spend to prove a winner, not production spend. Our $8.66 Studio annual is the render spend to ship three variants that then eat the $200–$500 ad-spend budget on Meta or TikTok.

The right framing for a brand running a monthly test:

line item	amount
Three variants rendered, retry-budgeted (our shelf, Studio annual)	$8.66
Ad spend to reach significance on three variants ($200/variant median)	$600
total monthly test cost	$608.66

Against a human UGC team at $25/variant render + $600 ad spend: $675. Against an agency at $500/variant render + $600 ad spend: $2,100. Our shelf is the render-spend line, not the ad-spend line — nobody disputes the ad-spend line, but the render-spend line is where the arithmetic diverges.

The lock-in

Product lock. The water bottle is a product reference image uploaded as @Element2. Kling O3 Pro's five-element cap gives us room for the presenter (@Element1), the product (@Element2), and three more slots (e.g. a brand logo, a packaging shot, a lifestyle backdrop) if the brief calls for them.

Presenter lock. The presenter's identity reads from a four-image reference set — front, 3/4, profile, full-body. The build amortizes to pennies across three variants and the next 20 tests on the same presenter.

Hook variable. The three hooks are prompt differences at the opening second, not reference differences. Variant A: "Here's the only bottle I'll drink from this year." Variant B: (mid-sip) "This is the thing that got me through today's workout." Variant C: (macro, voice-over) "Here's the bottle." Each prompt reads the same two references and renders a different opening.

Why three and not five

Pixelmotion's piece lists five-variant tests as common. The reason we ship three and not five is the credit math and the significance math meeting in the middle:

variants	render credits (Kling O3 Pro, 15s, voiced)	retry-budgeted	$ Studio annual rendered	$200/variant ad spend	monthly cost
2	900	1,050	$5.78	$400	$405.78
3	1,350	1,575	$8.66	$600	$608.66
5	2,250	2,625	$14.44	$1,000	$1,014.44

The render-spend scales linearly; the ad-spend scales linearly. The breakpoint is budget-driven, not credit-driven: on a $600/mo Meta budget, three variants fit; on $1,000, five do. Our Studio annual plan ships both.

A silent b-roll variant for the brand that does not want a voiced presenter

Some brands run product ads with no spoken voiceover — product-only visuals with a music bed and overlaid text. Kling O3 Pro silent bills 24 cr/s = 360 credits for 15s silent. Three silent variants: 1,080 credits = $5.94 Studio annual, retry-budgeted to $6.75.

That is the brand's choice; the credit meter does not force one or the other.

Where the keep-rate math breaks

The 83% keep-rate assumption holds for a locked presenter + locked product in frame. Three cases where it drops:

• Presenter holding the product in a complex grip (bottle rotation mid-sip, product being poured). Measured keep-rate falls to ~70% across 10 runs. Retry budget doubles. On three variants, expected total ≈ 1,890 credits = $10.40 Studio annual.
• Low-contrast product on a low-contrast backdrop (white bottle on a white table). Kling's reference reads the presenter cleanly but drifts on the product edge; keep ~75%. Workaround: swap backdrop in the prompt, not the reference. No credit cost to the workaround.
• Fast-cut hook with a camera-move opening second (dolly-in, orbit). Kling's motion signature on R2V holds through 15s but keep on reference drops ~5% in our tests. If the brief is motion-heavy, Seedance 2.5 at 70 cr/s 720p is the engine; three 15s variants bill 3,150 credits = $17.33 Studio annual — more than double Kling's rate for the motion-signature upgrade.

Where our shelf wins

Three variants in one plan. Studio annual at $82.50 effective ships three 15s variants with 13,625 credits of monthly headroom — enough for 30 more variants, or a different test every day.

Reference lock across variants. The presenter and the product hold identical across all three variants, which is the hard requirement for an A/B that tests the hook variable alone. Meta's ad test attribution is clean when the only variable is the hook, confounded when the presenter drifts between variants.

Silent/voiced toggle in-plan. Three silent variants bill 20% less than three voiced; the test can swap mid-plan at no extra subscription.

No render-queue tax. The editor stitches a lower-third or a logo endcard at no render cost; one competing agency ships the same three variants with per-export fees that double the per-variant cost.

Our 300-credit free account carries one Kling O3 Pro 10s voiced clip or roughly 12s silent, with 0 credits to spare on a single 15s voiced variant. Render one variant against your presenter and product references before you commit the three-variant pass — the identity-hold on both the presenter and the product is the thing to check, not the per-second headline.

The $200–$500 per-variant ad-spend floor is the number the SERP opens on. The $2.88 per-variant render-spend on Studio annual, retry-budgeted and reference-locked, is the one no listicle publishes — because it only exists on a shelf that bills a credit meter against one locked product and one locked presenter and five engines on the same plan.