A minute of generated picture costs $3 to $42, and the narration costs five cents

The comparison pages never divide by sixty

Four comparison pages, all read. Crazyrouter's table, dated 29 April 2026, runs six rows under columns headed Cost per 5s Video and Cost per Second, and the longest unit anywhere on the page is five seconds. Vidpros, dated 16 September 2025, does print a cost per minute. It gets there by dividing a monthly subscription by the number of videos it assumes you make: Google Veo 3 at $19.99 a month becomes an AI video cost per minute of $19.99, assuming single video creation. A monthly subscription divided by one video is a subscription price with a different label on it.

CostGoat, a calculator page that does chase the primary source, is the fair one: its per-second figures match Google's exactly, $0.05 for Lite and $0.40 for Veo 3.1. It still states no cost per finished minute. One of the four has the prices right and none of the four divides by sixty. Nobody buys a second of video, and nobody buys a monthly plan to make one video: the thing on the invoice is a finished cut. Nor does any of the four separate the picture from the audio, which is the split that explains why a pipeline and a single video model price differently.

A monthly subscription divided by one video is a subscription price with a different label on it.

Off the official pages, the same minute runs $3.00 to $42.00

Google's Gemini API pricing page prints Veo under a header reading Paid Tier, per second in USD: Veo 3.1 Standard with audio is $0.40 at 720p and 1080p and $0.60 at 4k, Fast is $0.10 at 720p, Lite is $0.05. OpenAI heads its own table Video generation models, prices per second, and lists sora-2 at $0.10 for 720p, then sora-2-pro at $0.30 for 720p, $0.50 at 1024p and $0.70 at 1080p. Runway prices in credits and states the rate on the same page: $0.01 per credit. gen4_turbo is 5 credits a second, gen4.5 is 12, aleph2 is 28 with a 56 credit minimum.

Multiply by sixty. Veo 3.1 Lite is $3.00 a finished minute, and Runway's gen4_turbo arrives at the same $3.00 from the other direction. Veo 3.1 with audio is $24.00. sora-2-pro at 1080p is $42.00. OpenAI's Batch tier halves its own column, so on that one page the cheapest minute is $3.00 and the dearest is $21.00, a factor of seven before a vendor has even been chosen.

Google's two price lists also disagree about what they are measuring. Vertex AI prints Veo 3.1 video plus audio at $0.40 for 1 count and never says on the page what one count is. The Gemini API page prints the identical $0.40 under a header that says per second. Same vendor, same number, one unit stated, which is why a per-second claim about Veo should cite the Gemini API page and not Vertex.

The Veo rate the comparison layer quotes was retired on 8 September 2025

Google announced the cut on its own developer blog, dated 8 September 2025, in two lines: Veo 3 now $0.40 a second, was $0.75; Veo 3 Fast now $0.15 a second, was $0.40. That is 46.7 per cent off the standard tier in one day: a minute of picture went from $45.00 to $24.00 overnight.

A page updated in July 2026 still opens by stating that Google's official Veo 3 API pricing on Vertex AI is $0.75 per second, and prices a standard 8 second clip at $6.00. At the rate both of Google's own pages carry, that clip is $3.20. The page overstates the number it is undercutting by 87.5 per cent, and advertises a saving of up to 97 per cent against it. The error runs in the direction that makes the discount look bigger.

The other direction is just as common. The Crazyrouter table gives Veo3 $0.10 a second where Google says $0.40, Sora $0.12 where OpenAI's 1080p row says $0.70, and Runway's Gen 4 Turbo $0.10 where Runway's own page says 5 credits, which is $0.05. Three rows checkable against a vendor's page, three contradictions. The page also asks to be trusted over the vendor in as many words: use Crazyrouter pricing and model metadata as the source of truth before publishing current prices or endpoint claims. It then sells access to those same providers at up to 30 per cent off.

The error runs in the direction that makes the discount look bigger.

Google's own list prices the audio inside a generation at $12.00 a minute

Vertex AI lists Veo 3.1 twice, and the gap between the two rows is the whole argument. Video plus audio generation, described as high quality video with synchronised speech and sound effects, is $0.40. Video generation, the same model with the audio removed, is $0.20. The difference is $0.20 a second, which is $12.00 for a finished minute of the audio Veo makes alongside the picture. At 4k the pair is $0.60 and $0.40, so the audio costs the same $12.00 there. Google's list charges the same for both halves: $12.00 of picture, $12.00 of speech and sound effects.

Google sells that minute of speech separately, priced per million tokens, and its own pricing page states the conversion: audio tokens correspond to 25 tokens per second, so sixty seconds is 1,500 tokens. Gemini 2.5 Flash Preview TTS charges $10.00 per million for audio output, which is $0.015 a minute. The $20.00 rows make it $0.030. ElevenLabs publishes an extra minute price of $0.17 to $0.20 across its paid plans and advertises low-latency speech as low as five cents a minute on Business. The audio inside a Veo generation is $12.00. A narrated read of the same minute is between one and a half cents and twenty cents.

Per-character prices need a denominator, and this one comes from the vendor: ElevenLabs lists 121,000 credits against roughly 121 minutes of text to speech at one credit a character, so the vendor's own page treats about 1,000 characters as a finished minute. Convert on that and the speech vendors disagree wildly. Runway sells eleven_multilingual_v2 at 1 credit per 50 characters, which is $200 per million characters, on the very page that sells the picture. OpenAI's tts-1 is $15.00 per million. Thirteen times apart, and set against a $24.00 minute of picture both are rounding error.

The audio track inside a Veo generation costs $12.00 a minute. Google will read you the same minute for one and a half cents.

Eight seconds is the unit, so a sixty second video bills as sixty four

Veo 3.1 caps a generation at 8 seconds, returns one video per request, and runs at 24fps. Sixty divided by eight is 7.5, so a sixty second video is at least eight generations and bills 64 seconds of output to deliver 60. At the standard rate that is $25.60, not $24.00. Extension exists, 7 seconds at a time up to 20 times, but extended output is 720p only, so above 720p eight separate full-price generations is the only route.

Sora's unit is different, and a per-second table hides it: up to 20 seconds a generation, six extensions, 120 seconds maximum, so the same minute is three generations.

Then the retakes, which no price page models. One discarded 8 second Veo Standard take is $3.20. One discarded sora-2-pro take at 1080p is $5.60. Eight shots with a single retake each is 128 seconds billed, $51.20 for one finished minute, and one retake a shot is optimistic. Nobody publishes how many attempts a usable shot takes. The one page that puts a number on it says to multiply base costs by 2.5x, and gives no method, no sample and no source, so there is no retake multiplier on the chart above.

Who pays for a failure is where the official pages disagree, and only two of them say anything. Runway states its rule under its own heading: moderated generations have the same credit cost as successful generations, and a moderated request reaches your code as a failed task, so a rejected take looks free in an error log and is not. Its task failures page adds that credits are not refunded for safety input failures, unlike other failures, and that you should not retry these generations. Google's Veo table carries the opposite sentence, scoped narrowly: inside a note about an audio processing issue it says you will only be charged if your video is successfully generated. Read literally that covers a request that produced no video, not one that produced a video you cannot use. OpenAI's guide lists failed as a job state and warns that copyrighted characters, copyrighted music and reference images with human faces are rejected, and says nothing about the charge.

What to do about it

1. Convert every quoted rate to one finished minute before comparing two vendors: per second times 60, and on Runway credits per second times $0.01 times 60.
2. Make the quote carry the picture and the read on separate lines. Google's own list charges $12.00 a minute for the audio inside a generation and one and a half cents to read the same minute at its text to speech rate.
3. Ask the vendor in writing what a moderated generation and a failed one cost, because only Runway publishes a rule and Google's is scoped to one failure mode.