Cost Per Finished Second: Normalizing AI Video Pricing

AI video cost per second sounds like a number you could just look up. It is not. As of August 25, 2026, the major video-generation vendors quote prices in four incompatible units — dollars per second of output, credits per second, dollars per million tokens, and per-generation or per-frame line items in the fine print — and none of them convert to another without a formula the vendor rarely states.

The stakes are practical, not academic. A team comparing Google’s $0.40-per-second Veo 3.1 Standard rate against Runway’s 12 credits per second, Kling’s 6–8 credits per second on a different credit system, and Seedance 2.5's price per million tokens is not comparing prices at all — it is comparing units. Get one conversion wrong and the whole budget model is wrong. And every published rate prices a generated second, while what a production actually buys is a finished one — the take you keep, after the takes you don’t.

This reference states one normalization formula, applies it to every model with a rate that can be sourced to a published page, and — just as deliberately — refuses to apply it where the arithmetic cannot be done honestly. Two of the most-discussed names in the field fall in that second bucket, and that finding is the strongest demonstration of the thesis this post has.

Start with the verbatim units, because the verbatim units are the whole problem. Google’s pricing page states Veo 3.1 is billed “per second of video” — the cleanest formulation in the field. Black Forest Labs and Lightricks publish the same shape for FLUX 3 Video and LTX-2.3. Runway prices every model in credits per second, with a separately published dollars-per-credit rate. Kling prices in credits on a different credit system with no first-party dollar conversion we could locate. And ByteDance’s Seedance 2.5 bills the generated video itself in tokens — dollars per million of them.

Every one of these shapes is defensible on its own terms. Per-token billing matches how a multimodal model actually meters its own compute; credits smooth out multi-model platforms; per-frame upscale pricing tracks the real cost driver of that operation. The problem is not any single unit — it is that a buyer comparing vendors has to do the unit conversion themselves, and most published “cost of AI video” comparisons silently don’t.

Normalization needs three terms: the vendor’s base rate converted to dollars per second, any mandatory upscale or enhance pass required to reach delivery quality, and a multiplier for the takes you generate but don’t keep. The first two come from published pricing. The third does not exist anywhere in published form — so it enters the formula as an explicit assumption, never as a measured figure.

One vendor term softens the multiplier slightly. Google states that Veo billing only fires on a successfully generated video — a partial mitigant, because it covers hard generation failures, not the far more common case of a generation that succeeds technically and still isn’t a usable take.

A boundary note on scope: within a single vendor, the cheapest path through a rate card is usually a draft-first workflow — iterate at the cheap tier, render the keeper at full quality. That within-vendor arithmetic is the subject of our draft-tier economics playbook, and this reference deliberately does not restate its per-vendor math. This post owns the layer that piece doesn’t: the conversion between vendors, where the units themselves disagree.

The table below holds every model and tier whose rate we could trace to a published page as of August 25, 2026. The unit column is verbatim — written the way the vendor states it, because paraphrasing units is how comparisons go wrong. The final column is the normalized base rate in dollars per second, before the keeper multiplier: derived cells state their conversion, and rates that reached us through aggregator relays rather than a first-party page load are labeled as exactly that.

* Base rate before the keeper multiplier. Derived cells state their conversion — Runway rows are credits per second × the published $0.01-per-credit API rate. ⚠ MiniMax’s rate card reached us through aggregator relays of the official card, not a first-party page load; spot-check before budgeting. † Runway’s Aleph 2 and hosted-Seedance figures rest on a single first-party fetch of Runway’s API pricing docs, not independently cross-confirmed.

One rate in this table has a story of its own: MiniMax’s $0.13-per-second 2K rate is the number behind the ad-economics analysis we published in July — cross-linked once here; this table only records the rate.

A normalization table earns its credibility by what it refuses to include. Two vendors are excluded from the column above — ByteDance, for its own Seedance 2.5 pricing, and Kuaishou, for Kling’s V3 and O3 tiers — not because they are obscure, but because no honest per-second dollar figure can be constructed for them from published, verifiable sources as of August 25, 2026.

The Seedance case deserves the emphasis, because it is the purest expression of this post’s thesis. Seedance 2.5 is one of the most capable video models in the field — our field comparison of FLUX 3, Seedance 2.5, and the Gemini video stack covers the capability side — and yet its cost per second is unknowable from public information. The token rate is published; the tokens-per-second-of-video relationship is not, at least not anywhere we could verify against a primary page. When we first covered this in early August, ByteDance had published no conversion; a formula has since begun circulating among aggregators, but its identical wording across supposedly independent blogs is a red flag, not a confirmation. Until the vendor publishes the conversion, the honest per-second price of Seedance 2.5 is: undefined.

Kling’s exclusion is procedural rather than structural — the model is credit-priced with subscription tiers (Free, $10, $37, $92, and $180 per month, all credit-based) and a relayed exchange rate, but its own pricing pages refused to load on direct fetch. A rate we cannot read from the vendor is a rate we do not put in the table.

Even among the normalizable vendors, three billing-shape wrinkles will corrupt a spreadsheet that assumes “rate × seconds” is the whole story. Each one is a published, first-party term — sourced to Runway’s API pricing docs and its consumer plans page — so the trap is not hidden pricing, it is a unit that doesn’t match the one you expected.

A fourth trap spans platforms rather than line items: the same model can carry different prices in different units depending on who bills you for it. Seedance 2.5 hosted on Runway is priced at 20 credits per output second plus 10 credits per second of reference video — $0.20 and $0.10 at Runway’s published $0.01-per-credit API rate — while ByteDance’s own direct rate for the same model is per million tokens, a unit that cannot be compared to Runway’s without the very conversion nobody has verifiably published. The platform door doesn’t just change the price; it can change the unit.

The formula’s third term — attempts per keeper — is where every “cost of AI video” analysis either gets honest or quietly makes things up. Here is the state of the evidence: none of the vendors in this reference (Google, Black Forest Labs, MiniMax, ByteDance, Kuaishou, Runway, Lightricks) publishes a first-party figure for what fraction of generations are usable on the first attempt. That is a gap in the field, not a research failure — and it constrains every vendor’s cost story equally.

The secondary literature does not rescue the number. Our research for this reference surfaced a claimed academic pass-rate figure attributed to a recent arXiv paper that could not be verified against the paper’s own abstract or an attempted full-text fetch, and a scatter of failure-rate percentages attributed to unnamed “published benchmarks” with no named study, no date, and no sample size reachable. None of those figures appear in this post, in either direction — a number we cannot trace is a number we do not print, even with a hedge attached.

The second table this reference contributes is about the sources themselves. For each vendor: does a first-party dollars-per-second rate exist, is a credit-to-dollar conversion published, is a token-to-second conversion published, and is an upscale term stated? “Not located” means exactly that — not locatable in this research as of August 25, 2026, which is a different claim from “does not exist.”

Read as a trend, the census splits the field into two postures. Google, BFL, and Lightricks treat a stated per-second price as a competitive asset — LTX in particular publishes the cleanest first-party rate in the field, with no intermediary unit at all. The credit and token vendors sit behind at least one conversion layer, and in Kling’s and ByteDance’s cases, behind a layer this research could not verify first-party. Pricing opacity is not random; it correlates with platform business models that benefit from customers thinking in the platform’s own currency.

Three worked examples, all at the illustrative default of A = 3 attempts per keeper. LTX-2.3 Fast at $0.04/s: $0.04 × 3 = $0.12 per finished second. Veo 3.1 Fast at 720p: $0.10 × 3 = $0.30 per finished second. Runway Gen-4.5 with a 4K upscale on the keeper: ($0.12 × 3) + $0.288 = $0.648 per finished second — where the upscale pass, applied once at the top of the per-frame range, costs 2.4× a single generation attempt. And the floor trap compounds: three 1-second Aleph 2 attempts bill the $0.56 minimum each, so a single kept second costs $1.68 — six times the headline $0.28/s rate.

Two practical notes for teams operating this at scale. First, the keeper multiplier dominates everything at these rates — cutting A from 3 to 2 saves more than switching between most adjacent vendors, which is why prompt discipline and draft-tier iteration matter more than rate-shopping. Second, video APIs burn budgets fast enough that cap behavior matters: our census of what actually happens at an AI spend cap applies with full force to a video pipeline mid-batch. For the plumbing itself — wiring these APIs into an agentic creative pipeline with retries and budget guards — our video-API integration guide covers the architecture.

Looking forward, expect the unit problem to get worse before it gets better. As multimodal models unify text, image, and video generation in one architecture, token billing becomes the natural internal meter — Seedance 2.5 is likely the leading edge of that shape, not an outlier. The normalization burden will keep shifting onto buyers until a vendor decides that a published, verifiable cost-per-finished-second becomes a competitive weapon — the same dynamic that eventually forced per-request cloud pricing into calculators and rate cards. Teams that build the conversion discipline now, with the keeper assumption stated in the open, will be the ones whose video budgets survive contact with production. If you want that discipline built into your creative or media operation — model selection, budget guards, and the pipeline around them — that is the work of our AI transformation engagements, and for video-heavy ad accounts specifically, our paid media practice runs exactly this arithmetic before a campaign brief is approved.