Kineo AI Video Index — October 2026

Key findings

What the renders say this month

Customer renders · October 2026

Engines measured on customer renders

Time from request to finished film, length of the finished video, reliability among renders that actually started, and the automatic coherence score — customer renders only. Kineo’s own price per finished video for every engine is kept on one page, not repeated here: Seedance, Kling and Veo in one place.

Shown separately, never added

Kineo internal test renders, indicative

Engines without enough customer renders in this window. Kineo’s own accounts test every engine before and after a change; those films are excluded from every customer number above and shown here only so each engine still has a measured render time and length. No rates are published for them.

“indicative” = too few renders to rank engines; read it as an observation, not a benchmark. A dash means there was nothing to measure in the window.

Provider cost

What the raw video costs at the provider

The per-second list price the inference provider (fal.ai) publishes for each model, at the resolution Kineo renders and without audio — the raw clip, before the script, narration, retries and assembly a finished film needs. Checked October 5, 2026 and October 6, 2026.

Methodology

How the index is measured

• Window: 30 days, September 7, 2026 – October 6, 2026 (UTC), read from Kineo’s production database on October 7, 2026. The queries are versioned with the numbers, so every edition is computed the same way.
• Who is measured: the customer table uses customer renders only; house accounts (the founder’s and test accounts) are kept apart and never added in. An engine without enough customer renders in the window (customer sample n < 10) appears only under “Kineo internal test renders, indicative”: time to film and length, no rates. $0 dry runs are excluded everywhere.
• Kineo 1 (stock-footage assembly) is not a generative engine and is not covered.
• Sample sizes are given as ranges, never as counts: n ≥ 100, n 10–99, n < 10 — indicative (an observation, not a benchmark). This edition — Seedance 1.5, customer renders: n ≥ 100; Kling 2.5, Veo 3.1, Kling 3, MiniMax H3, Seedance 2.5 and Omni Flash, Kineo internal test renders: n < 10 — indicative.
• Time to film: from the first server record of the request to the moment the finished MP4 was saved to the library, matched by the generation id. When nobody keeps the page open, a background check that runs every five minutes saves the film, which can add up to about five minutes.
• Length: the duration of the finished MP4 as saved at delivery; it matches a reading of the MP4 header within one second on every film where both were recorded.
• Reliability: finished films ÷ (finished films + renders that failed with a recorded error), counting only renders that actually started (at least one scene sent to the model). Left out, and why: requests stopped before any scene was rendered — account checks (plan, quota, credit balance), Kineo’s script and length checks, and server errors before the first scene — because no engine render happened; and renders that ended without a recorded outcome, because an unknown outcome is not a failure. When those unknown outcomes exceed 20% of an engine’s started renders, the table says “not enough data” instead of a rate.
• Coherence: an automatic judge (a text model, version five of Kineo’s rubric) rates every delivered film 0–100 on two questions — does the narration tell what the person asked for, and do the planned scenes show what is being said — and averages them; 75+ reads as coherent, under 50 as off. It reads text (the request, the narration, the scene plan), not pixels.
• Provider price: the per-second list price the inference provider (fal.ai) publishes for the model and resolution Kineo renders, without audio, with the date it was checked — a raw clip price, not the cost of a finished film. Kineo’s own price per finished video is not repeated here: it is on “Seedance, Kling and Veo in one place” (usekineo.com/seedance-kling-veo-in-one-place), read from the same source that charges.
• Aggregates only — no individual creator, prompt or account appears anywhere.