Affiliate disclosure: we may earn commissions when you sign up through some links below, at no extra cost to you. This never affects our pricing data, comparisons, or recommendations. Learn more.

AI Pricing Guru Labs · Direct API measurement

CogVideoX-3 API pricing: what each accepted video second actually cost

Z.ai lists CogVideoX-3 at $0.20 per video. After one billed request became unretrievable and four sampled seconds failed our usability rule, the measured cost rose to $0.0516 per accepted second—29.0% above the nominal five-second rate.

8 direct API requests 2 matched prompts Text + image input Official pricing checked 2026-09-17

Buyer verdict

Budget about 29% above the clean per-second math

The simple calculation is $0.20 ÷ 5 seconds = $0.0400 per second. The end-to-end run debited $1.60 for eight requests and yielded 31 accepted seconds, raising the observed rate to $0.0516. That is the more useful planning number for this tested workflow, not a universal CogVideoX-3 quality estimate.

Request success

87.5%

7 of 8 retrievable

Accepted seconds

31/40

77.5% end to end

Median generation

96.2s

Successful calls only

Effective cost

$0.0516

Per accepted second

Shared zero-based scale

Nominal price versus measured accepted-footage cost

Both bars use the same $0.00–$0.06 axis. The effective bar includes the billed but unretrievable request and rejected sampled seconds.

Nominal five-second math$0.0400
Measured cost per accepted second$0.0516
$0.00$0.03$0.06 per second

At this measured rate, 1,000 accepted seconds would cost about $51.61, versus $40.00 from list-price arithmetic alone.

Text versus image input

Image-to-video followed the prompts more reliably

ModeRetrievableAccepted secondsPrompt adherenceCost / accepted second
Text-to-video3/415/1533.3%$0.0533
Image-to-video4/416/2068.8%$0.0500
Do not overgeneralize. This was two product-ad prompts with two attempts per fixture/mode. The result supports a bounded buying insight—reference frames helped with exact text and directed motion here—not a universal ranking of CogVideoX-3 input modes.

Methodology

How we decided whether a second was usable

Fixture and grading contract

  • • Two matched product prompts: a NOVA bottle orbit and AURA headphones dolly.
  • • Four requirements per clip: stable product, specified object motion, camera move, and exact full-clip text overlay.
  • • One frame sampled at the midpoint of each requested second.
  • • A second was rejected for a severe geometry, continuity, duplication, or text artifact that made the footage unusable.

Limits

  • • Midpoint samples can miss brief inter-frame artifacts.
  • • One request was billed but could not be retrieved after its submission response was lost.
  • • Results apply to quality mode, 1280×720, 30 fps, five seconds, and no audio.
  • • Provider billing came from authenticated bonus-balance readings before and after the run.

Choose the workflow, not just the sticker price

Compare a multi-model video API before committing to one direct provider

This benchmark was run and billed through Z.ai. Check Novita's current image and video API catalog if you want a commercial multi-model alternative, then rerun the same acceptance test on the exact model you select. Novita was not the provider measured here, and its catalog should not be assumed to reproduce CogVideoX-3. AI Pricing Guru may earn a commission at no extra cost to you; it did not influence the benchmark.

For Z.ai text-model rates and the separate coding subscription, see our Z.ai pricing guide. The Coding Plan should not be treated as CogVideoX-3 API credit.

Sources and freshness

Official price and API behavior checked . Data owner: AI Pricing Guru Data. Re-verification cadence: daily. Official sources are monitored for price, duration, request-schema, and result-schema changes.