Ross Mike burned through $24,000 of AI model usage last week.
He paid about $400 for it. Two $200 subscriptions. That's the whole bill.
That number is the real headline of Greg Isenberg's new episode on GPT-6 Astra, and it's why I'm breaking it down today. Everyone else is posting 3D train demos. Greg and Ross spent 23 minutes on the only question that matters: how do you actually make money with GPT-6 Astra while the pricing is this heavily subsidized?
Their answer has three parts: two prompts you should run on your existing app tonight, nine prompts for building income on top of the model, and one genuinely wild story about a $561 DIY smart speaker that I haven't stopped thinking about.
Here's the full episode:
What GPT-6 Astra Actually Is (Minus the Demos)
Quick context, because most Astra coverage is useless.
GPT-6 Astra is OpenAI's most expensive model, priced roughly on par with Anthropic's Fable. But two things change the math:
→ If you're on a Codex subscription, usage is heavily subsidized. Hence Ross's $24K of compute for a couple hundred bucks.
→ It's efficient. It completes tasks in fewer steps, so the high per-token price doesn't sting the way you'd expect.
Ross put it bluntly: "The saddest thing people can do with a model like GPT-6 Astra is generate landing pages."
He's right. We could do that two generations ago. The interesting stuff is what this model unlocks that nothing before it could touch.
Run These Two Prompts on Your App Tonight
Before any new business ideas, Ross gave two prompts for founders who already have a product live. Both take minutes to kick off.
1. The performance update. Tell the agent: "I want to run a performance update. My pages and API calls aren't as fast as they need to be. Review my entire app and suggest a performance update."
Ross ran this on one of his apps. API speeds went from 800 milliseconds to 20-30 milliseconds.
Not 20-30 percent faster. 800ms down to 20ms.
2. The security audit. Ross has a small payments product with real users. He ran a security audit through Astra and, in his words, the model was basically like "my boy, you would have been cooked if I didn't read all this."
He says this is something the previous generation of models couldn't reliably do. If strangers are logging into something you built, run it this week.
One bonus tip from the episode: when you ask Astra for UI, give it a reference image. Ross fed it the old Windows XP network icon and got a design he actually shipped. Without a reference, "it kind of does too much."
The 9 Money Prompts (Greg Brockman Retweeted These)
Greg Isenberg tweeted a list of nine Astra prompts and OpenAI's president retweeted it. Here's the list, with the ones I'd prioritize marked.
1. The bill renegotiator. "Go through my internet, phone, and software bills, jump into each provider's chat support, and negotiate them down or cancel what I'm not using." Uses Astra's computer-use ability. Pays for the subscription by itself.
2. Agency into software. This is the one I'd steal first. "Pick one service business in a niche. Reverse engineer the exact workflow they sell to clients. Break it into steps, tools, inputs, outputs, human judgment points, and places where the work gets slow or expensive. Then design the simplest AI product that could replace the first version of that service and charge $500 to $5,000 a month."
That's a full product discovery process in one prompt.
3. The marketplace watcher. "Watch Facebook Marketplace and Craigslist in my city for cameras, furniture, and bikes listed way under market, and text me the second one's mispriced." Greg pitched this for young people who can't land a job right now and want a few hundred a day flipping. Ross screenshotted it for his church group chat.
4. The one-person company dashboard. "Look at my docs, notes, Stripe exports, analytics, customer calls, and project list. Build a weekly operator dashboard: what's making money, what's wasting time, what customers are asking for, what I should stop doing, and the three highest-use actions for next week. Be blunt and show your work."
If you run a solo SaaS, this is your fractional COO for the price of a prompt.
5. The agent-opportunity audit. "Look at how this business works and find the tasks we should give to agents before hiring another person." Greg said this is now his default filter before any hire: should I just give this to an agent? For each task it estimates human time, cost of mistakes, difficulty of automating, and the first safe version to deploy, prioritizing anything that saves money or creates revenue within 30 days.
6. The browser operator. Give Astra a workflow and let it click through real sites, collect data, fill forms, and note what broke. Then it hands you a repeatable SOP and an automation plan. Ross used this to comb through Canada's tax agency website and find things Google couldn't.
7. The nightly QA team. "Every night, open my app on a real phone, go through signup, checkout, and the main flows, and screenshot anything broken or confusing." The keyword is confusing. UX bugs count.
8. The competitor spy. "Sign up to my top three competitors, sit inside their product and their emails, and send me a monthly report on every new feature, price change, and the thing they do better than me." Monthly, not daily. Enough to stay sharp without obsessing.
9. The viral lead-capture game. Build a browser game around one mechanic, with progression, scoring, and one reason someone would send it to a friend. Then wire in email and SMS capture tied to your core product. Greg's point: most AI-built games are cute demos nobody shares. Tie one to your product and it turns into a distribution channel.
The $561 AI Speaker: From Idea to Merged PR in 30 Minutes
Now the part of the episode that actually moved me.
Ross has wanted a Jarvis-style home device since he was a kid. He owns a HomePod and loves the hardware ("the best looking speaker, in my opinion") but Siri is terrible. Jailbreaking it? Not possible. He checked.
So he asked Astra a different question: what if I build my own?
Here's the sequence, and remember this is a webdev, not a hardware guy:
→ He told the agent he had Blender installed and asked it to connect. It did, and modeled a HomePod-style speaker.
→ He asked how he'd load his personal agent (named Ruth) onto physical hardware, and pasted his repo. Astra came back with a Raspberry Pi build, a wiring diagram, and a full parts list.
→ Budget estimate: $350 to $450 before tax. Actual Amazon.ca cart: $561. He bought it on the spot.
→ When he panicked ("I don't know nothing about hardware"), he asked Astra to lay out the physical setup in Blender. It rendered the exact arrangement: Pi, speaker, connectors, voice recognizer.
→ Then it wrote the integration code for his agent repo and opened a pull request. He merged it. The code is waiting for the parts to arrive.
→ In a separate thread, it even surfaced Chinese suppliers who could manufacture the speaker shell for a real product run.
Idea, 3D model, parts list, purchase, wiring diagram, merged PR. Half an hour of conversation.
He's also planning to wipe a dust-collecting PS4 and turn it into a Linux home lab because Astra told him he could. His line stuck with me: "Intelligence has gone up, but the bravery for people to try new things hasn't gone up."
Vibe Manufacturing Is the 2024 Moment, Again
Greg's framing for all this: remember when Lovable, Bolt, Replit, and v0 hit in 2024, and suddenly anyone could vibe code a web app?
The same shift is happening right now for physical products. Greg's calling it vibe manufacturing.
The barrier to a niche hardware product used to be: learn CAD, learn circuits, find a manufacturer, pray. Now the model does the 3D file, the bill of materials, the code, and the supplier list, and you do the judgment calls.
Greg's prediction: thousands of $1M-a-year businesses come out of this wave. Niche physical products that solve one real pain, cheap to prototype, sold with Meta ads driving traffic. Markets too small for a hardware giant to care about and exactly the right size for one founder.
Will Anthropic and everyone else catch up on this kind of capability? Obviously. Ross, a self-declared regular of the Claude ecosystem, says this is the first time he'd pick the other subscription. That race is good for all of us. But Astra ushered this particular era in, and the subsidized window won't last. Both of them said it plainly: take advantage before they charge the real price.
What I'd Actually Do With This Episode
If I were at $5K-$20K MRR reading this, here's my order of operations.
This weekend: run the performance prompt and the security audit on your existing app. Two prompts and an evening of runtime, and you've pulled real risk out of your product. No excuse.
Next week: run prompt #5, the agent-opportunity audit, before your next hire. Map every task in your business and score it for automation. If you want a deeper framework for that mapping, this guide on business automations pairs well with the prompt.
This month: pick one service business you already understand and run prompt #2 against it. The agency-into-software prompt is the closest thing I've seen to a repeatable SaaS idea generator that starts from proven demand instead of a brainstorm.
And if you want the pattern behind this playbook: I did the same breakdown when Fable 5 dropped, and Greg's local AI episode got the same treatment in my local models breakdown. Every model release creates a 6-to-24-month window where the tools are ahead of the crowd. The founders who win each window are just the ones who run the prompts while everyone else watches demos.
FAQ
What is GPT-6 Astra?
GPT-6 Astra is OpenAI's newest flagship model, and its most expensive, priced roughly in line with Anthropic's Fable. It stands out for computer use (operating a real browser and desktop), hardware and 3D workflows (driving tools like Blender, generating wiring diagrams and parts lists), and deep codebase work like security audits. Codex subscribers currently get heavily subsidized usage.
Can you make money with GPT-6 Astra without being a developer?
Yes, several of the nine prompts in Greg's list need zero code: the bill renegotiator, the marketplace mispricing watcher, and the competitor spy all run on Astra's computer-use ability. Ross's hardware build is the bigger point though. He's a web developer with no hardware background, and the model carried the parts he didn't know. The gap that matters now is bravery, not skill.
How much does GPT-6 Astra cost?
On raw API pricing it's OpenAI's most expensive model, comparable to Fable. In practice, a Codex subscription (the $200 tier in Ross's case) subsidizes it massively: Ross measured over $24,000 of equivalent model usage in one week across a couple of $200 subscriptions. Both Greg and Ross expect that subsidy to shrink, which is their argument for building aggressively now.
What is vibe manufacturing?
Vibe manufacturing is Greg Isenberg's term for using AI models to take a physical product from idea to prototype without traditional hardware skills: the model handles 3D modeling, the bill of materials, firmware code, and even supplier sourcing. It's the hardware equivalent of 2024's vibe coding wave, and Greg predicts thousands of $1M-a-year niche product businesses will come out of it.
Steal the Full Playbooks, Numbers Included
Every week on the Profitable Founder Podcast I sit down with bootstrapped founders making $100K to $10M a year and get the real numbers behind what they built and which channels actually brought the revenue. If Greg's prompts got your gears turning, the interviews are where you see them turned into MRR.
