AI Side Hustles Mostly accurate, with one big caveat

By the Income Reality Check Editorial Team · Updated 2026-06-18 · 7 min read

Claude Fable 5 made an entire YouTube video — the income math it skips

Verdict: Mostly accurate, with one big caveat. The video really was built by AI from a single prompt, and the model claims hold up — but making a video and making money from one are two different jobs.

Nate Herk’s video “Claude Fable 5 Made This Entire Video By Itself” does something most make-money clips don’t: it tells on itself. The avatar isn’t filmed, the voice is a clone, the script was written by Claude, and the whole thing — research, voiceover, motion graphics, editing — came out of one /goal prompt while Nate was at the gym. No dollar figure is promised. But the implied pitch is obvious to anyone watching: if AI can produce a finished YouTube video this good, you could run a channel on autopilot. That inference is where the reality check lives.

What the video actually claims

The headline is literal. Nate says he typed one prompt into Claude Code, walked away, and came back to a fully edited video he never saw a frame of while it rendered. Claude read Anthropic’s announcement, fact-checked it, wrote the script in his voice, sent it to an ElevenLabs voice clone in sub-minute chunks (longer generations drift), rendered each chunk on a HeyGen Avatar 5 model, stitched the clips with FFmpeg, and built every motion graphic as animated HTML code timed to the words. Then it reviewed its own output frame by frame and re-rendered anything that looked off.

He’s also specific about the engine. Anthropic released Claude Fable 5 — described in the video as a “mythos class” model, the tier above Opus — and Nate frames it as the reason the video can exist at all. He quotes the launch numbers: a 50-million-line Ruby migration done in a day, beating Pokémon Fire Red on raw screenshots, $10 per million input tokens and $50 per million output.

To his credit, Nate doesn’t oversell the repeatability. He says outright that copying his exact prompt probably won’t reproduce his result, because he’s got “hyperframe skills” already built into his setup. He says he could likely do the same thing with Sonnet now that he’s built the workflow once. And he flags the cost: this one experiment ate roughly 40% of his $200-a-month Max plan in a single hour.

Do the Fable 5 claims actually hold up?

Mostly, yes — which is worth saying plainly, because the genre is full of invented capabilities. Anthropic’s own model documentation lists Claude Fable 5 (claude-fable-5) as its “most capable widely released model,” generally available from June 9, 2026, priced at exactly $10 / $50 per million tokens — matching Nate’s numbers. Claude Mythos 5 is the tier above it, offered in limited availability through Anthropic’s invitation-only Project Glasswing, which lines up with his “locked to vetted security partners” description. The pricing he quotes isn’t hype; it’s the rate card.

One small correction for calibration. That $10 / $50 rate is the Fable tier’s price, and it’s the same rate the older Opus 4.7 carried. The current flagship Opus 4.8 actually runs cheaper — $5 input, $25 output per million tokens — so Fable 5 is the premium option, not a discount. For a workflow that burned roughly 380,000–400,000 tokens, the model bill alone is real money before any subscription enters the picture.

What the method actually requires

Here’s the part the smooth one-prompt framing glides past: the prompt sits on top of a paid stack and a pile of pre-built assets. Strip the demo down and you’re paying for at least four things at once.

Run the math on a single four-minute video. HeyGen’s Creator plan ships 200 credits; at about 20 credits per minute on the newest avatar models, four minutes is roughly 80 credits — more than a third of your monthly allowance on one clip. Add the slice of the $200 Max plan, the voice-clone subscription, and the fact that the “skills” doing the editing took real work to build before any of this ran hands-free. The “one prompt” is true. The “$0 and no skills” version of it isn’t on offer here, and Nate doesn’t claim it is.

That’s the production side. It’s genuinely impressive, and largely solved. The trouble starts at distribution.

Who actually wins this game

The people who win with a workflow like this aren’t beginners typing a prompt for the first time. They’re operators who already have the inputs the demo quietly assumes: a voice clone trained on years of their own footage, an avatar trained on their face, a custom skills library, and an existing channel with an audience and a track record. Nate has all four. That’s why his output looks like a real Nate Herk video and not generic AI filler — the human fingerprint was baked in long before the prompt ran.

Strip those away and you get a different result. A first-timer can absolutely generate a polished four-minute clip. What they can’t generate is the thing that makes it earn: an audience that’s already watching, and a format distinct enough that YouTube treats it as a creator’s work rather than a template.

What you’d realistically earn

Start with the platform gate. To earn ad revenue at all, you need into the YouTube Partner Program: 1,000 subscribers and 4,000 valid public watch hours in the past 12 months (or 10 million Shorts views in 90 days), per YouTube’s own requirements. Most channels never clear that bar, and the ones that do don’t do it in a week.

Then the rate. NerdWallet pegs a conservative ad-revenue estimate at $1 to $5 per 1,000 views, after YouTube takes its 45% cut. So a video that lands 10,000 views — already a strong number for a new faceless channel — might return $10 to $50. To approach the $5,000-a-month figure these automation channels imply, you’d need on the order of a million-plus monthly monetized views, sustained, in a decent-paying niche. NerdWallet’s framing is blunt about it: YouTube “is not a get-rich-quick platform.”

Now the part that should stop anyone planning a video factory. YouTube’s monetization policy explicitly bars “inauthentic content” — mass-produced or repetitive uploads — and calls out, by name, “AI-generated content made with generic templates giving the impression of mass production” without the creator’s original insight. In July 2025 YouTube renamed its old “repetitious content” rule to “inauthentic content” and began evaluating whole channels, not single videos. By January 2026 the enforcement was real: reporting on the crackdown describes 16 channels with a combined 35 million subscribers and 4.7 billion lifetime views terminated under that policy. A pipeline where every upload “looks, sounds, and moves the same” is precisely the risk profile.

The nuance matters, though, and cuts in the creator’s favor. YouTube hasn’t banned AI. Properly disclosed, AI-assisted content with genuine editorial input keeps full monetization. Nate’s single experimental video — his voice, his style, real information, one-off — is nowhere near the line. A hundred identical clones of it, published daily by someone with no audience, would be.

The disclosure question (U.S. readers)

One more cost that doesn’t show up in the token count. If you use a cloned voice or an AI avatar to promote anything, U.S. rules apply. The Federal Trade Commission has moved on exactly this: its 2024 rulemaking on AI impersonation targets synthetic likenesses and voices, and existing FTC endorsement guidance requires clear, conspicuous disclosure when AI generates or materially alters promotional content. Outside the U.S., labeling rules differ — the U.K.’s ASA and the EU’s AI Act transparency provisions land in similar territory — but the principle travels: if a synthetic person is selling something, say so.

Who this is (and isn’t) for

This makes sense for an established creator who already has the audience, the trained voice and avatar, and a reason to cut production time from days to an hour. If that’s you, Fable 5 plus a skills library is a genuine force multiplier, and the cost per video is a rounding error against your existing revenue. It does not make sense as a from-zero passive-income plan. If you’re hoping to point a prompt at YouTube and collect checks while you sleep, you’re buying the one part of the business AI has actually automated — production — and skipping the two it hasn’t: building an audience and staying inside the monetization rules.

For more on that second job, see our breakdown of how to beat the new YouTube algorithm in 2026 for faceless channels, and our companion look at what Claude Fable 5 really did when it “built a $10k website in minutes”.

What to remember

The video is honest about the thing it demonstrates and quiet about the thing it implies. AI can now produce a broadcast-quality video from one prompt — that’s real, the model is real, and the price is on Anthropic’s rate card. What no prompt does for you is earn the 4,000 watch hours, build the audience, or keep your channel on the right side of YouTube’s inauthentic-content rules. Production got cheap. The business didn’t.