AI Side Hustles Half-true — works only if you do the unspoken work

By the Income Reality Check Editorial Team · Updated 2026-07-01 · 7 min read

GLM 5.2 vs Claude: the “free AI beats Claude” pitch and the money hiding in the gap

Verdict: Half-true — works only if you do the unspoken work. The model really is cheaper and often better on routine tasks, but the money the video hints at sits behind exactly the scarce engineering skill it describes.

Nate B Jones, on his channel AI News & Strategy Daily, opens with a claim that sounds like a free lunch: GLM 5.2 is an open-source model that’s dirt cheap (free if you host it yourself), and on the fat middle of everyday work — brochure sites, first-pass copy, routine coding — it’s not just good enough, it’s often better than Claude. So why hasn’t every company dumped its expensive frontier-model contract? That’s the real subject of the video, and buried inside his answer is a pitch aimed straight at you: there’s a “golden goose moment” here for anyone who can build the plumbing companies can’t. Is that opportunity real? Mostly yes — with a catch most viewers will miss.

What the video actually claims

Jones makes two linked arguments. The first is technical: GLM 5.2 is roughly “98% cheaper” than Claude and matches or beats it on what he calls “center of distribution” work — common tasks with familiar shapes and outputs a human can check fast. He’s careful not to trash Claude; his point is that most knowledge work is center-of-distribution, so a model that nails it deserves to be taken seriously.

The second argument is where the money lives. He says companies can’t just swap a model name and route everything to the cheap option, because “you’re not replacing a model call, you’re replacing a whole work system.” A model, in his words, is a “brain in a jar” — useless without a harness (the prompts, memory, tool-calls and routing logic wrapped around it). Rewriting that harness for an open-source model is hard technical work, and the talent to do it is so scarce it “can charge anything it wants.” His conclusion, aimed at entrepreneurs and agencies: “This is a golden goose moment. You can really go to town and basically promise to save people a ton of money on tokens.”

He cites Flo Crivello’s company Lindy, which publicly moved off Claude, as proof the savings are real. He’s honest that Lindy “had to essentially rewrite their harness from scratch.” Hold onto that admission.

What the method actually requires

Start with the price gap, because “98% cheaper” is doing a lot of work. On the international API, GLM 5.2 lists around $1.40 per million input tokens and $4.40 per million output. Anthropic’s published rates put Claude Opus 4.8 at $5 input and $25 output per million tokens, with Sonnet-class models at $2–$3 input and $10–$15 output (Anthropic pricing docs). So against Opus you’re looking at roughly 70–82% cheaper per token — real, but not 98%. You only reach “free” or near-zero by self-hosting the open weights on your own GPUs, and that swaps a token bill for a hardware-and-ops bill.

Now the part the excitement skips. Lindy’s own write-up of its DeepSeek migration is blunt: “Changing a model name was easy. Proving that users would still trust the assistant took the work.” That work was offline evaluations, prompt re-optimization, staged rollouts to internal users first, live monitoring of retention metrics, and gradual traffic ramping (Lindy blog). Inference costs on migrated traffic fell about 90% — but only after a specialist team earned the right to move that traffic at all.

That’s the unspoken work, and it isn’t marketing or SEO this time. It’s senior AI engineering. Here’s the rough shape of what a real switch involves:

CNBC reported this exact shift in June 2026: buyers are moving from “tokenmaxxing” to efficiency, and startups are switching to cheaper models to survive (CNBC). The demand is genuine. What the video glosses is that capturing it is a skilled-services business, not a plug-and-play side hustle.

Who actually wins this game?

Three groups, and you should know which one you’re in. First, AI-native product companies like Lindy, where every cent of inference is margin — they have both the incentive and the in-house engineers. Second, agencies and consultants with real ML talent who can sell the refactor as ROI (“we’ll cut your token bill 80%, here’s the eval to prove quality held”). Third, the frontier labs themselves, who keep winning because switching is hard.

That last point matters. Anthropic’s revenue run rate reportedly climbed past $47 billion in 2026, with roughly 80% coming from enterprises and Claude Code alone at a $2.5 billion annualized clip (CNBC). A cheaper competitor exists and the incumbent is still growing like a weed. Why? Because the harness, the team habits, and the tool integrations lock customers in. The friction Jones describes is the same friction that makes his “golden goose” real — and the same friction that keeps most would-be goose-farmers out.

What you’d realistically earn

If you’re a strong AI engineer or you run an agency with one, the opportunity is legitimate and well-paid. U.S. AI-specialist salaries in 2026 commonly run $145,000–$310,000 base, and independent LLM specialists command steep day rates because demand outruns supply. The Bureau of Labor Statistics projects software-developer employment to grow 15% from 2024 to 2034, far above the 3% average across occupations (BLS). A consultant who can reliably deliver a model migration that holds quality can bill four to five figures per engagement.

If you’re the 22-to-45-year-old aspiring entrepreneur watching this video hoping to “save companies money on tokens” without that background? Your realistic near-term earnings from this specific play are closer to zero. Not because the market isn’t there, but because you’d be selling a rewrite you can’t yet perform. The honest on-ramp is months of learning to build agent harnesses, evals, and routers first — the skill is the product. The token savings are just what you sell.

There’s no passive version of this. None.

Who this is (and isn’t) for

This makes sense if you already write code, understand how LLM tool-calls and context windows work, and can commit real hours to building and testing evaluation pipelines — ideally with a client or two who’ll pay for outcomes. It also fits agency owners who can hire or partner with that talent and package it as a service. It does not make sense as a quick income idea if you’re non-technical, expecting AI to “do it all for you,” or hoping to monetize in weeks. The video’s own example (Lindy) needed a specialist team and a staged rollout to pull it off. Take that as the floor, not the exception.

What to remember

Jones is more honest than most creators in this niche — he says out loud that a cheap model without a harness is a brain in a jar, and that building the harness is hard. That candor is why the verdict is “half-true” rather than “hype.” GLM 5.2’s price and quality are real. The business opportunity is real. What’s missing from the 95,000-plus viewers’ takeaway is that the opportunity and the difficulty are the same thing: the money exists because the work is hard, and it flows to the people who can already do it. If that’s you, sharpen the saw. If it isn’t yet, the first project is learning, not billing. For the broader question of building an AI-driven business, see our looks at starting a one-person business with AI in 30 days and how to use AI in your business in 2026.