Same task. Same quality bar.
Opus 4.8 charged $2.38. GLM 5.2 charged 44 cents.
That's the whole reason this open-source model went viral on Twitter. People are calling it the ChatGPT moment for local AI, and Greg Isenberg brought his friend Amir on the podcast to show how he actually runs it inside Cursor and Codex.
Amir is the guy Greg calls first whenever he's testing new AI stuff. He's also mid-way through a mentality shift that I think most of us are about to go through: from "I don't care what tokens cost, I'm just vibe spending" to "wait, why is my usage limit gone by lunch?"
I watched the full episode and pulled out the numbers, the setup, and the one trick (model chaining) that makes a cheap model punch at frontier level.
Here's the episode:
What GLM 5.2 actually is
GLM 5.2 is an open-source model from Z.AI, the Chinese lab behind the GLM family.
Open source means you can download the weights and run it on your own machine. No API bill, and no waiting for Anthropic to un-throttle you when you hit a limit.
The headline numbers from the episode:
- 1 million token context window
- 81 points on Terminal Bench 2.1, about 4 points behind Opus 4.8
- 62.1% on the long-horizon task eval, versus 69.2% for Opus 4.8
So no, it's not better than Opus. It's about 90% of Opus.
But 90% of Opus at roughly one fifth of the price changes how you build. That gap is the entire story.
Greg admitted on camera that benchmarks go in one eye and out the other for him, and Amir agreed. His test is simpler: build something real with it and see how it feels next to the model you already use. He rebuilt the hero section of one of his apps with it, then had it turn a wall of feature screenshots into a carousel and a Bento grid. One prompt each. It followed instructions in a way local models simply didn't a year ago.
The math: 44 cents vs $2.38
Amir mapped out a typical coding task: about 50,000 input tokens and 85,000 output tokens.
→ GLM 5.2 through OpenRouter: $0.44
→ Opus 4.8: $2.38
Call it 5x cheaper for near-frontier output on execution tasks.
On a single prompt, who cares. Two bucks is nothing.
But nobody runs a single prompt anymore. You run agents overnight. You retry. You let a coding agent loop on a bug for 40 minutes. I've hit my Claude limit on day one of a billing cycle, and so has Amir ("when Fable came out, I hit my limit the first day").
At that volume, 5x is the difference between a $200 month and a $1,000 month. Per seat.
How to set it up (two ways, 10 minutes)
This is the part nobody was showing before this episode. Two routes, both credit-based. You don't need to buy any hardware.
Route 1: Cursor + Z.AI directly
- Get an API key from Z.AI (they make GLM 5.2)
- In Cursor settings, paste it into the OpenAI key field
- Override the OpenAI endpoint with Z.AI's endpoint
- Add "GLM 5.2" as a custom model
Done. Cursor now calls GLM 5.2 like any other model.
Route 2: Codex + OpenRouter
- Grab an OpenRouter key and load $20 of credits
- Create a profile in Codex with the model name, endpoint, and context window
- Switch to GLM 5.2 from the CLI whenever you want the cheap model
Amir's advice for anyone starting: skip the hardware debate entirely, load $20 into OpenRouter, and start testing today. OpenRouter runs the open models in the cloud and charges per token, so you get the open-source pricing without owning a single GPU.
Model chaining: the trick that makes it work
GLM 5.2 has a real weakness: no vision. It can't look at a screenshot of your app.
Amir's workaround is the smartest thing in the episode.
He feeds screenshots to Opus 4.8 and asks it to describe the layout in detail: "tell me what you see on the hero section and lay it out." Then he switches to GLM 5.2, hands it that written layout, and lets it make the actual code changes.
Expensive model thinks. Cheap model types.
OpenRouter calls this fusion models. Amir calls it model chaining. Either way, the pattern is: use the frontier model for the 10% of work that needs judgment or vision, and route the other 90% (the token-heavy execution grind) to the model that costs a fifth as much.
You need a harness that's model-agnostic for this, which is exactly why he runs everything through Cursor, Codex, or Claude Code instead of locking into one provider's app. Same logic I covered in the local AI business ideas piece: the founders who win the next cycle are the ones whose stack survives any single provider having a bad week.
The token subsidy is ending
This is the part of the episode that actually matters for your P&L.
Greg's analogy: Uber. VC money subsidized your rides until you were hooked, then prices climbed. AI labs are doing the same thing with tokens right now. You're building your whole workflow on top of subsidized compute.
Amir sees the corporate side of this first hand. Companies that spent the last year "token maxing" as an AI adoption metric are now canceling direct Claude API access because the bill got scary. Satya Nadella has started talking about human capital plus token usage as a combined cost line at Microsoft.
And the governance stories are almost funny. Amir is literally getting paid to tell companies things like: John in marketing should probably not use a frontier thinking model to format an email.
His conclusion flipped from a year ago: stop token maxing. Token minimize, output max.
If you can show $200 in and $1,000 out, fine, keep spending. But the subsidy runs out sooner or later, and the founders who already know how to route work to cheap models will keep their margins when it does.
Should you buy a Mac Studio? (No.)
Every time a local model trends, Twitter fills up with people telling you to drop $5,000 on a Mac Studio.
Greg asked Amir straight: do people actually need local hardware?
His answer: no. GLM 5.2 is resource-hungry, and most consumer machines can't run it well anyway. "If I just have the right machine, I'll get the result" is not how it works. OpenRouter gives you the same model, in the cloud, for $20 of credits.
The one nuance: if you're running massive token volume every single day, the math on owning hardware starts to work, because GLM 5.3 and 5.5 are coming and memory prices aren't dropping. Buying the machine now is a bet that future models make your fixed cost look cheap.
For everyone at $5K to $50K MRR: that's a later problem. Credits first, hardware maybe never.
What I'd actually do with this
My take, as someone whose AI bill is a real line item every month.
1. Set up OpenRouter this week. $20, ten minutes. Even if you never switch your daily driver, you now have a fallback for the next time your main provider throttles you.
2. Route your grunt work. Refactors, test generation, CSS changes, boilerplate. That's GLM 5.2 territory at 44 cents a task. Keep Opus or Fable for architecture decisions and anything involving screenshots.
3. Track cost per outcome, not tokens. Amir's ROI framing is right. $200 in, $1,000 out is a business. $200 in, vibes out is a subscription you'll cancel in six months.
4. Watch this space for arbitrage. A 5x cost drop on execution-grade AI makes services profitable that weren't profitable last quarter. I wrote about this exact pattern in AI arbitrage businesses: when your input cost falls 80% and your client's willingness to pay doesn't, that gap is your margin.
The bootstrappers who treat model choice like any other unit economics decision are going to quietly out-margin everyone still defaulting to the most expensive model for every prompt.
FAQ
Is GLM 5.2 better than Claude Opus 4.8?
No. On the benchmarks cited in the episode, GLM 5.2 scores 81 on Terminal Bench 2.1 versus Opus 4.8 about 4 points higher, and 62.1% versus 69.2% on long-horizon tasks. It's roughly 90% of Opus quality at about one fifth of the cost, which makes it a fit for execution tasks, not for replacing your frontier model everywhere.
Can GLM 5.2 see images or screenshots?
No, it has no vision capabilities. The workaround from the episode: have a vision-capable model like Opus 4.8 describe the screenshot in text, then hand that description to GLM 5.2 to act on. You get frontier-level results on front-end work at the cheap model's price.
Do I need special hardware to run GLM 5.2?
No. It's too resource-intensive for most consumer machines anyway. The recommended path is OpenRouter: load $20 of credits and call GLM 5.2 in the cloud from Cursor, Codex, or Claude Code. Only consider buying hardware if you're running heavy token volume daily and want to lock in fixed costs before future models arrive.
How much cheaper is GLM 5.2 than Opus 4.8?
About 5x on a typical coding task. Amir's example: 50,000 input tokens plus 85,000 output tokens cost $0.44 on GLM 5.2 through OpenRouter versus $2.38 on Opus 4.8. The gap compounds fast once you're running agents and long tasks daily.
What is model chaining?
Using two or more models in sequence on the same task: an expensive thinking model for planning and vision, a cheap execution model for the actual changes. OpenRouter calls it fusion models. It needs a model-agnostic harness like Cursor or Codex, and it's the main way founders in the episode get frontier quality at open-source prices.
Every week on the Profitable Founder Podcast, I sit down with bootstrapped founders doing $100K to $10M a year and pull apart exactly how they run leaner than you do. The token bill conversation is coming up more and more.
