What a second of AI video costs, by channel

RenderDrama records published per-second rates for generated video. Sixteen channels publish one: the lowest is a cent a second on a speed-tuned variant, the highest $0.70 at 1080p. Each figure carries the page it came from and the date it was read. As of 2026-09-22.

Inclusion rule. Channels that publish a per-second rate, or a credit rate with a credit price. A channel selling only monthly allowances is not converted. Order. Ascending by the lowest published rate, ties alphabetical.

1Three billing words, defined

Three words that decide what a second actually costs have entries of their own: credit billing, resolution tiers, prompt extension. Every figure on them is quoted from a vendor page and carries the date it was taken.

2What is recorded here

One page per channel, each answering the same questions: what a second costs, what the figure covers, how long a single generation can run, and what the vendor leaves unstated. Rates come from the vendor's own public pages. Where a vendor bills in credits, the conversion is shown along with the assumption it rests on rather than presented as a plain number.

What is kept out matters as much. Monthly plans are not divided into per-second figures, because the result depends on how much of an allowance gets used. Ranges stay ranges. And the three costs that actually decide a production's bill, the usable-take rate, the clip length and whether failures are billed, are given their own places rather than folded into the headline.

3Start here

• The ledger — sixteen channels side by side.
• The ratings — sixteen tools on seven published criteria.
• How scoring works — what each criterion asks, and what leaves it blank.
• Which tools join the segments — the most polarised of the seven criteria.
• Which tools read a script — twelve of sixteen publish nothing about it.
• SceneMixer — credits per second with a published conversion.
• MiniMax — two model lines that meet at $0.08 a second.
• Vidu — the flattest resolution curve, and a half-price off-peak rate.
• Kling AI — a rate that at least doubles with each resolution step.
• Runway — credits with the credit price on the same page.
• LTX Studio — a four-to-one band rather than a rate.
• Veo — three tiers eight times apart, and a published answer on failed generations.
• Sora — a batch tier that halves every figure on the page.
• Wan — a doubling ladder, with the clip ceiling printed beside it.
• PixVerse — credits per second, and one dollar reference for every credit.
• Dreamina — a per-second figure that belongs to one plan.
• WaveSpeed AI — a cent a second, with a condition on every figure.
• Runware — a list price and a discount printed in the same cell.
• fal — the one page here that names its own billing unit.
• Novita AI — sound priced as half again per second.
• Replicate — two video rates beside a run-time meter.
• Luma — credits per second, with the credit price missing.
• Pika — credits attached to features rather than to seconds.
• Morph Studio — allowances counted in videos, not seconds.
• invideo AI — per-seat plans and approximate credit counts.
• CapCut — a price the help centre says to read inside the app.
• HeyGen — clip length, rollover and concurrency published; no rate.
• Synthesia — minutes per plan, which is the rarer missing half.
• Magnific — credits against named durations, valid for a year.
• Segmind — a per-second figure that prices the GPU instead.
• Hedra — allowances only, and a rollover question left open.
• Clip limits — how long one generation can run, where it is published.
• Unspent credits — what becomes of a credit you did not spend, and the seventeen channels that do not say.
• Parallel generations — how many a channel runs at once, and the one that sells the extra sessions.
• Method — what counts as a rate and what is left out.
• Alternatives, priced per second — the tools people switch between, read against the second.

4Costed questions

Eleven pages that take one question about what generated video costs, answer it from the figures already in the ledger, and show the arithmetic. No page adds an assumption the vendors did not publish.

• The cheapest second anyone publishes — the floor is down to a cent.
• What a credit is worth — the same second, two correct dollar figures.
• The costs outside the rate — compositing, regeneration, the clip ceiling.
• Prices that depend on when you generate — one channel halves off-peak.
• What the jump to a higher resolution costs — one and a half to five times, per channel.
• What a minute costs — under $1 to $25, arithmetic only.
• Which channels publish a rate — and why a monthly plan is not one.
• How long a single generation can be — the field most often missing from a price page.
• What happens to credits nobody spent — six of the vendors, five with a published answer.
• How many generations run at once — and who sells the extra sessions.
• One family, four published seconds — the spread runs fourteenfold.

5Explainers

Cost mechanics, written for somebody about to commit a budget. No vendor figure lives on these pages: every number that belongs to a product sits in the ledger with its source and the day it was read.

• Per-second billing — why the headline rate is the smallest part of what a project costs.
• Budgeting a series — a worked method for costing an episode before you generate anything.
• Costing a season — nine hundred seconds of picture at four rates, and the five costs beside them.
• Token billing — the third billing shape, where the same clip triples between output tiers.

6How entries are kept

A page in this ledger is dated by its content rather than by its build, so regenerating the site moves nothing. Where a seller publishes a figure different from the one recorded here, the seller is the record and the row is amended at the next pass.

RenderDrama records list prices published by vendors. They are not quotes, and nothing here is a recommendation to buy.