AI presenter video API: seven checks, vendor pages read Oct 2026

Before you pick an AI presenter video API, check seven things: whether it is live or rendered, how it counts length, what its API access includes, how it bills, how many jobs run at once, what resolution it outputs, and what happens to a failed job. The four vendors below publish enough to answer some of these; where a page is silent, the honest answer is that you do not know yet.

Everything about a vendor comes from its own page, read 2026-10-06: Synthesia, Tavus, Vozo and Sync Labs with its docs. Sume's column comes from Models, Generate avatar video and Generation admission.

What do the pages say, check by check?

What do the empty cells tell you?

Failure handling is blank for every vendor because none of the pages read states it. That is a question for a sales call or a test: submit a bad input and see whether you are charged. On Sume it is documented: the reservation taken at admit is refunded when a job fails, and a 402 is returned before any provider work if the balance cannot cover the reserve.

Resolution is the check that most often surprises teams. Sume's Avatar 1.0 video is 720p at this time, and H3 Max lip sync goes to 1080p. If a channel needs Full HD presenter footage, Synthesia's published Full HD downloads are a plain advantage that you should weigh against the rest.

How should you weigh them?

Run one real script through two vendors before you decide, and keep the evidence: the output file, the invoice line and the time from submit to result. Two vendors that look alike on a pricing page often differ most on revision cost, and revisions are where presenter budgets go.

• Live or rendered first. If the avatar must answer a person in real time, only the live products qualify, and Sume is not one.
• Then unit of length. A 60-second cap per job is fine for ads and explainers and a nuisance for a ten-minute lecture.
• Then billing shape against your month: a plan with minutes you will not use is more expensive than it looks.
• Then concurrency. A launch that needs 40 clips in an hour will queue on a small plan anywhere.

What does Sume want from you?

A ready avatar handle and a script, or ordered video_inputs with silence beats; a quality of standard, plus or max; and an Idempotency-Key. The estimated duration must land between 4 and 60 seconds. For first-frame approval before a full render, use the preview endpoints. For the checklist applied to a vendor with no public price at all, see the five-question vendor check, and for a first runnable job see the Python walkthrough.

Two Sume details are worth knowing before a trial. A multi-scene job uses one resolved avatar for the whole video, and scene backgrounds must resolve to one shared scene, so a script that needs three different presenters is three jobs (Generate avatar video, read 2026-10-06). A silence beat is a scene with a required duration and no script, which lets you leave a pause for a slide or a product shot without paying for filler speech.

Captions are optional and applied after generation, onto the clean final MP4, using the same script or scene text. Sume never burns captions into preview stills. If you want to approve the composition first, Avatar video previews make the first-frame stage separate from the full render, so you review the stills before paying for the talking video. Write down the preview-to-final step in your trial notes, because it is the closest thing to a revision control that a presenter API offers.