ToolScore

Video generation: ranked by the cost of one usable 5-second clip

The easy pick by value, not by hype. Our metric is what a usable 5-second clip costs — generated is not usable: motion artifacts, broken physics and dead audio get clips rejected, and every reject burns the same credits, so assume ×2–4 on the sticker for real work. Every $/clip below comes from official pricing pages (credit conversions marked); where the vendor hides the price the cell stays “—” until it can be verified. The Artificial Analysis Video Elo appears only as a quality reference.

The board just reshuffled — two names vanished
Sora is discontinued: web/app went dark 2026-04-26, the API on 2026-09-24, and stored data will be “permanently deleted” per the official help page. Veo is gone too: Veo 3.1 Preview shut down 2026-10-22 — its successor, Gemini Omni 1.2 Flash, now tops the paid table below. The lesson this hub is built around: rented video endpoints carry expiration dates, and the local table is the insurance.

Open-source & local — you pay electricity

No per-unit price exists here: the cost model is watts. A 170 W card draws ≈$0.03/hour at the US average rate ($0.18/kWh — the same default as our rent-vs-own calculator), and a rejected clip costs a re-roll, not a credit. Our own 3060 box runs a closed video pipeline (agent writes, Kokoro voices, ComfyUI renders) — but no verified 3060 video-model report exists yet, so no row below claims one. Sorted by value per VRAM — license traps cap the score.

Model License VRAM / needs Score
Wan 2.2 TI2V-5B 5B params Value pick Apache 2.0 24 GB (official) · <9 min/720p clip 88/100
LTX-2.5 19B params Only local audio+video Apache (code) · Community weights — · fp8/fp4 quants 76/100
HunyuanVideo 1.5 8.3B params Tencent Community · NC in EU/UK/KR 14 GB min (offload) 58/100
Mochi 1 Apache 2.0 ~60 GB · <20 GB in ComfyUI 52/100
Wan 2.2 T2V/I2V-A14B 14B params Apache 2.0 80 GB (official) 48/100
LongCat-Video 13.6B params MIT — 44/100
Kandinsky 6.0 Pro MIT · gated — 40/100
Open-Sora Apache 2.0 — 36/100
MiniMax-H3 33B params Community · US-banned — 6/100

Click a model for the full picture — license terms, hardware paths and the score breakdown on its review page. “—” in VRAM = verify pending: we publish hardware requirements only when a source confirms them. Scores are our editorial estimate (0–100) of value per VRAM. The 12 GB GGUF path is community territory — no verified 3060 report exists; everything else cites LICENSE files and model cards.

×2–4 rejectsthe multiplier on every price above

Pricing pages bill per generated clip; your keeper rate decides the real cost. Video is the medium where prompts miss most — at a 1-in-4 keeper rate, Gemini Omni’s $0.35 becomes $1.40 and FLUX 3 Draft’s $0.30 becomes $1.20 per usable clip. Local rows re-roll for electricity — ≈$0.03/hour on a 3060-class card — which is why the hit rate matters more the more you pay per click.
When paid wins — and when local wins
Paid: client-ready 1080p+ with audio yesterday and no GPU in the loop — Gemini Omni and FLUX 3, billed per project. Local: volume, iteration and control — batch clip pipelines where a rejected take costs watts, not credits (Wan 2.2 5B officially fits 24 GB; GGUF quants stretch it toward 12 GB). Our own closed pipeline: One GPU Beats Ten Subscriptions and the Local GPU AI leaderboard.

Pipeline economics — local electricity vs API vs premium subscriptions for a full script→voice→video loop — get their own article: Our Video Pipeline Runs Local; this hub sticks strictly to per-clip prices.

Quality reference — AA Video Leaderboard Elo

Blind-vote preference Elo from Artificial Analysis — read it next to the price, never instead of it. The curiosity worth pausing on: the Elo king is Wan 3.0 at 1175 — the same family whose Wan 2.2 release runs locally under Apache 2.0 (table 1); FLUX 3 follows at 1137.

Model Elo AA-listed price
Wan 3.0 1175 —
FLUX 3 1137 —

Snapshot 2026-11-12. Only these two rows were extracted from the board — no more rows are invented. AA-listed prices across the video leaderboard run $2.40–34.12/min; per-model attribution wasn’t captured (verify pending). Paid-table prices come from official pricing pages (Google AI, bfl.ai, Runway, Pika); Elo from the AA Video Leaderboard. Licenses from LICENSE files and model cards.

The stack around video generation

Local video gen is the top layer of a stack you can own: One GPU Beats Ten Subscriptions shows the closed pipeline our 3060 box actually runs (agent writes, Kokoro voices, ComfyUI renders — zero per-call bills for the media layer), and the Local GPU AI leaderboard ranks the hardware it runs on. The voice layer lives in the TTS hub, stills and thumbnails in the Image generation hub; for the thinking layer, see the AI Coding Plans ranking. The dedicated video deep dives — per-keeper math, our local pipeline, the fine-print traps — are in the Deep dives section below.

What to look for in a video generator

A clip you keep is the unit

Every vendor quotes the happy path per generated clip. Divide by your real keeper rate before comparing — ×2–4 is normal in video — and a local model that re-rolls for electricity changes the economics entirely.

Endpoints have expiration dates

Sora vanished in two steps (web 2026-04-26, API 2026-09-24) and took users’ data with it; Google sunset Veo 3.1 Preview a month later. Never route client work through a single rented endpoint without a fallback — the local table is that fallback.

Credits that burn monthly

Luma credits expire every billing cycle; Pika’s plan credits don’t roll over; Runway grants rollover only on Max (1 month) — while purchased credits survive (Pika 365 days, Runway unlimited). A subscription punishes every idle month.

License territories and free-tier rights

MiniMax-H3 bans the USA outright; HunyuanVideo 1.5 bans the EU/UK/Korea but not the USA; Pika’s free tier has no commercial license while Runway grants commercial rights on every plan. And Google stamps SynthID on every Omni video — a term, not a footnote.

Deep dives

Data freshness
Snapshot 2026-11-12. Video is the fastest-moving AI market on this site — entire products disappear (Sora, Veo 3.1 Preview) and successors inherit their prices overnight. Every number above carries its source, and cells without one stay “—” until verified.