brain/
conceptai-video-generation

AI video quality versus cost

Notes

AI video quality versus cost

One-line summary: As of 2026-08-18 the frontier Elo cluster (Omni Flash / MiniMax H3 / Seedance) is also among the cheaper APIs; the expensive Western names (Veo 3.1 Standard, Sora 2 Pro) are not the quality leaders on the independent board.

The insight

"Best" and "cheapest" used to pull in opposite directions. The 2026-08-18 fetch shows the opposite at the top: gemini-omni-flash is both #1 T2V Elo and ~$0.10/s official; minimax-h3 is tied on Elo at $7.80/min; veo-3-1 Standard is $0.40/s for mid-pack Elo. Paying more now often buys brand, 4K, or a control surface — not higher blind preference.

Evidence

Design implications

Default stack from this source: Omni Flash to iterate, Seedance 2.0 to animate a still, Wan3.0 when you need 30 s or a document, HeyGen/Act-Two when the job is a talking head or a performance take. Do not buy Veo Standard or Sora 2 Pro for "best quality" on current Elo.

Contradictions / tensions

  • SCOPE.md still parks "commercial-pricing / market-share" as out of scope; this concept exists because the queue question asked for cost-effectiveness.
  • HeyGen's pricing-page FAQ (3 / 20 cr/min) disagrees with the dedicated rate card (4–48 cr/min).
  • Elo is not linear utility; a 150-point gap is not "4× worse."

Open questions

  • What does a usable clip actually cost after retries / failed gens? OpenAI's pricing page charges per second requested; retry rates are uncited.
  • Local H3 electricity vs $7.80/min API — no official VRAM floor.

Related

Referenced by