Cost of 1,000 five-second AI avatar replies: Fabric vs H3 Max lip sync

Rendering 1,000 five-second replies on Sume's H3 Max lip sync costs $312.50 at 480p, $500 at 768p and $1,000 at 1080p, before text-to-speech. Fabric at 720p reserves $0.94 per 5-second clip, which is $937.50 for the same bank at the unrounded rate. Pre-rendering a bank is the practical alternative to a live avatar for questions you can predict.

The per-clip math

H3 Max lip sync bills ceil(duration_seconds) times the fal list rate per second times the 1.25 house margin. fal lists $0.05, $0.08 and $0.16 per second at 480p, 768p and 1080p, so Sume's rates are $0.0625, $0.10 and $0.20 per second. A 5-second clip is therefore $0.3125, $0.50 and $1.00.

Fabric's docs give a 5-second reserve of $0.94 at 720p, which is 5 x $0.15 x 1.25 = $0.9375. Multiply either by 1,000 for the bank.

What the bank does not include

Text-to-speech is billed separately by character, and your first batch will want a few retries. A failed job refunds its reservation, but a take you dislike is a successful job you still pay for, so budget for rerolls. Costs above are before any cent rounding on individual jobs.

Where this beats a live avatar

A bank pays once per answer, however many people watch it. A live session pays for every viewer's minutes. If 80% of questions repeat, render those answers once and keep a live agent for the rest. Use the 5-second floor deliberately: H3 Max lip sync rejects audio under 5 seconds, so shorter replies go to Fabric.

See the 10-second cost ladder for the same comparison at a different length, and fal's own list at the vendor page for the underlying rates.