AI video price per minute: normalize per-second and per-token prices

To compare AI video API prices, convert every price to dollars per finished minute at one resolution, then check what the clip needs beyond that, such as audio and length. Hedra's best-models page, read 2026-10-02, lists Omni Flash at $6.00 per minute at 1080p and Kling 3.0/O3 Pro at $20.16 per minute at 1080p Pro, which is a 3.4 times spread before you consider quality.

Divide a per-minute price by 60 to get dollars per second; multiply a per-second price by 60 to go the other way. Sume's own prices are read from its catalog, not from this post.

What do Hedra's numbers look like per second?

The per-second column below is simple division of Hedra's listed per-minute prices. Hedra lists the Seedance 2.0 4K price as approximate, so its row is too.

How do I handle per-token prices?

Some catalogs price per token. Sume's Video generation docs show a pricing_skus entry of per-1000-video-tokens for seedance-2 in the example response. To turn that into a clip price you need the token count for your resolution and duration, which the docs example does not give. Use the MCP dry_run preview or the usage.cost from a short test clip instead of guessing.

What else changes the real price?

Check resolution, clip length limits, audio, and whether failed jobs bill. Sume's Video Router docs state list price times 1.25 for catalog models, so compare on the delivered price, not the list price.

How do I keep the comparison honest?

Use the same duration, resolution, and audio setting in every row, date the table, and refresh it. Cost of 20 fifteen-second ads at per-minute prices applies the conversion to a batch.

Put the conversion in a spreadsheet with the source address and the date beside each price. When a price is approximate, as in Hedra's Seedance row, keep the tilde. Do the comparison on the same line items: video only, same resolution, and same audio setting, since a bundled price and an unbundled one are not comparable.