PILLAR GUIDE · SCRIPT → VOICE → VIDEO
The complete AI stack for YouTube
A YouTube video is three products stacked: a script, a voice and pictures. Each layer has a different best tool, a wildly different price, and a different place where quality is won. This is the full 2026 toolchain, layer by layer, ending in one number: what a professional AI video actually costs.
| Layer | Best tools | Cost per 8-min video | Where quality is decided |
|---|---|---|---|
| Script | Claude / ChatGPT / Gemini | $0.05–0.70 | Your prompt, structure & rewrite |
| Voice | Inworld / ElevenLabs / OpenAI TTS | $0.10–1.80 | Voice choice & pacing marks |
| Video | Seedance + Veo 3.1 Fast / Sora 2 | $30–40 | Tier routing & retake control |
| Total | The full stack | ≈ $32–43 | — |
Rates shown at each model's base tier, verified July 2026 from official vendor pricing pages and documentation. Vendors change prices without notice — see methodology.
Layer 1 — Video: route shots, don't pick one model
The video layer eats 95% of the budget, so it gets the most engineering. The mistake beginners make is picking "the best model" and generating everything on it; the professional pattern is routing shots by job:
At a typical 40 shots (~240 seconds of kept footage), this routing lands at $30–40 per video including realistic retakes — the full arithmetic is in the YouTube budget breakdown. Two disciplines protect that number: draft every prompt on the cheapest tier before rendering final quality, and track your retake ratio per shot type so the routing improves monthly. What decides perceived quality isn't the average shot — it's the open, the payoff and the thumbnail frame. Spend there.
Layer 2 — Voice: the cheapest layer, the most underrated
Here's the asymmetry nobody exploits: voiceover is under 5% of the budget but carries 50% of retention. Viewers forgive soft B-roll; they close the tab on a robotic voice. And yet this layer costs almost nothing if you buy it right. The market in mid-2026, converted to one honest metric — price per million characters (≈ 18–24 hours of speech):
| Provider | $ / 1M chars | Cost per 8-min video | Notes |
|---|---|---|---|
| Inworld TTS-1.5 Max | $10 | ≈ $0.07 | Ranked #1 on quality in 2026 — the value disruptor |
| Hume Octave 2 | $7.60 | ≈ $0.05 | Cheapest; strong emotional adaptation |
| OpenAI TTS | $15 / $30 HD | ≈ $0.10–0.21 | Solid, easy API, no subscription needed |
| ElevenLabs Starter | $5/mo · 30k credits | ≈ $1.17 (at plan limit) | Commercial rights start here — never monetize free-tier audio |
| ElevenLabs Creator | $22/mo · ~100 min | ≈ $1.80 | Professional voice cloning; the realism benchmark |
| Google / Amazon standard | $4 | ≈ $0.03 | Audibly synthetic — fine for drafts only |
The honest recommendation: start on Inworld or OpenAI TTS — at $0.07–0.10 per video the cost is a rounding error and the quality is production-grade. Upgrade to ElevenLabs Creator only for two specific reasons: professional voice cloning (a consistent branded voice is a real channel asset) or its still-unmatched expressiveness on long narration. Three rules regardless of provider: check the license tier before monetizing (ElevenLabs' free tier explicitly forbids it), write for the ear — the voice can only read what the script gives it — and regenerate flat sentences individually rather than re-rendering whole scripts.
Two adjacent audio decisions: music comes from AI generators like Suno (subscriptions around $10/month with commercial rights on paid tiers — verify current terms) or classic royalty-free libraries; and native model audio (Veo, Sora) covers diegetic sound in the shots themselves, not narration — the split is explained in native audio vs voiceover.
Layer 3 — Script: where the video is actually won or lost
Uncomfortable truth first: the script layer costs the least and matters the most. Retention curves are written here — before a single second is generated. The three frontier assistants (Claude, ChatGPT, Gemini, each ~$20/month on consumer plans, or cents per script via API) are all capable scriptwriters, and the differences between them are smaller than the difference between a lazy prompt and a good one. What separates scripts that retain from scripts that don't:
- Feed it research, don't ask it to know. A model writing from your source material (transcripts of top videos in the niche, your notes, real data) produces specifics; a model writing from memory produces generic filler viewers can smell.
- Structure beats prose. Prompt for the skeleton explicitly: a 15-second cold open that states the payoff, an open loop every 60–90 seconds, a mid-video re-hook, no summary outro (endings kill session time — point to the next video instead).
- Write for the voice layer. Short sentences. Spoken rhythm. Read the draft aloud once — anything you stumble on, the TTS stumbles on worse.
- The rewrite is the job. Treat the AI draft as a first draft from a fast junior writer: keep the structure, replace the clichés, inject your actual take. Channels that publish unedited AI scripts converge on the same voice — and the algorithm buries sameness.
- One workflow trick that compounds: keep a running "style document" (your best hooks, phrases, banned words) and paste it into every script prompt. The model gets more you every month.
Script cost per video: effectively free on a subscription you likely already have, or $0.05–0.70 via API depending on model and draft count. It's the only layer where spending more money buys nothing — only spending more attention does.
The assembled pipeline, start to finish
- Research & script (LLM + your editing) — 60–90 min, ≈ $0.20
- Voiceover (TTS, regenerate weak lines) — 15 min, ≈ $0.10–1.80
- Shot list from script (one prompt per narration beat) — 20 min, free
- Draft visuals (Seedance 480p / Vidu, kill weak prompts cheap) — 45 min, ≈ $4
- Final renders (routed: Seedance 720p bulk, Veo Fast / Sora 2 heroes) — 60 min, ≈ $28–36
- Edit (cut to the voice track, captions in the editor — never generated in-frame) — 90–120 min
- Shorts pass (re-cut 2–3 vertical moments; Sora 2 portrait for gaps) — 30 min, ≈ $2–4
Total: ≈ $32–43 and 5–6 working hours per video. That number is the whole strategic story of AI YouTube in 2026: production cost has collapsed to near-zero relative to ad revenue (break-even around 11–14k views at a $3 RPM, far less in high-RPM niches — see the revenue math), which means production is no longer the moat. Everyone can make cheap video now. The remaining moats are niche selection, script quality and iteration speed — all three of which this stack frees your hours for.
Frequently asked questions
Can I run this entire stack for free to start?
Nearly. Google's $300 Cloud credit covers the video layer for your first month or two (Veo 3.1 Lite), free LLM tiers write drafts, and TTS free tiers cover testing — but note voiceover free tiers (including ElevenLabs') typically exclude commercial rights, so budget the ~$5–6 entry tier before monetizing.
Should I use one subscription that does everything instead?
All-in-one "faceless video" apps trade cost and control for convenience — you'll pay 3–10× the per-video cost of this stack and inherit their visual style, which the algorithm learns to recognize. The stack takes one extra hour to learn and pays that back weekly.
What should I upgrade first as revenue comes in?
In order: a cloned branded voice (ElevenLabs Creator — audio identity compounds), then premium hero-shot budget (Sora 2 Pro / Veo Standard for opens and thumbnails), then editing help. The script layer never needs money — only your hours.
Run your own numbers. The cost calculator applies your clip length, resolution and a realistic retake buffer across every model at once.