GPT-6 Astra: How I’d Make Money With It

The most interesting thing about GPT-6 Astra isn’t another impressive game demo. It’s what happens when you give it a problem that actually matters to your business.

Make an existing product faster. Find security problems before they become expensive. Turn a service workflow into software. Take a hardware idea you’ve been sitting on and work out the parts, the layout, and the code.

That’s where my brain goes: How does this save money, create revenue, or make something possible that felt out of reach?

I compared notes with Ras Mic on the practical use cases we’re excited about. I brought nine prompts for running and growing a business. He brought a personal AI speaker project that went from an idea to a parts order, a Blender layout, and a merged pull request in about half an hour.

Not a finished speaker in half an hour. But a serious path toward building one. That distinction matters, and the progress is still pretty wild.

Start With the Product You Already Have

Before inventing a new business, use Astra to improve something people already use.

Ras described Astra as OpenAI’s most expensive model at the time of our conversation, but also unusually efficient at completing tasks in fewer steps. His argument wasn’t that intelligence is cheap in absolute terms. It was that the value can be excellent when a stronger model needs less back-and-forth to finish meaningful work, especially through a subsidized Codex subscription.

Two tasks stood out immediately.

Run a performance audit

If you have a web app, mobile app, or existing codebase, ask the agent to review performance across the whole application rather than optimize one isolated function.

Review this application for performance bottlenecks. Investigate slow pages and API calls, explain what is causing the delays, and propose a prioritized performance update.

Ras tried this on one of his apps and reported that response times went from roughly 800 milliseconds to 20 to 30 milliseconds.

That’s one project’s result, not a promise for every application. But it’s a much more compelling use of a powerful model than generating another landing page.

Run a security audit

Ras also had a niche payments product with real people using it. When he asked Astra to audit the code, it identified security issues he considered serious.

If your app handles payments or customer data, this is the kind of work that deserves attention now, not after you’ve shipped five more features.

Audit this application for security risks. Explain the findings, identify the most urgent issues, and propose fixes for review.

Treat the output as an input to your security process, not proof that the app is secure. The useful point is that you can put a capable agent to work finding problems you may have missed.

Give UI work a visual reference

Ras’s best design results came from giving Astra a reference image. For one interface, he used the old Windows XP network icon as inspiration. Very specific. Very nostalgic. Very over 30.

The result had a clearer visual direction. Without a reference, he found that the model tended to do too much.

The lesson is simple: don’t just ask for something that looks good. Show the agent what kind of good you mean.

Nine Astra Prompts for Saving Money and Building a Business

I shared these nine use cases publicly, and Greg Brockman reposted them. What I like about the list is that it moves beyond generating content into operating workflows.

Some are personal productivity tools. Some could become products. Others help a founder make better decisions with the information already sitting inside the business.

1. Renegotiate recurring bills

Internet, phone, and software bills are easy to ignore because each one looks manageable. Together, they become a meaningful expense.

Review my internet, phone, and software bills. Identify unused services and opportunities to reduce costs. Use each provider’s support chat to explore better rates, and flag proposed changes or cancellations for approval.

The opportunity is straightforward: give an agent a clearly defined cost-reduction job. This is especially appealing when getting a better rate means navigating a website, finding the right support channel, and spending time in a chat queue.

2. Turn a service business into software

I’m obsessed with the question of which agency workflows can become software products.

Don’t start by asking for a generic AI startup idea. Start with something a business already pays another business to do.

Choose a service business in a specific niche and reverse-engineer what it delivers. Map the steps, tools, inputs, outputs, human judgment, and expensive bottlenecks. Then design the simplest AI product that could deliver an initial version of that service at $500 to $5,000 per month.

That price range is a product-design target, not validated demand. The output won’t be perfect. What it gives you is a concrete workflow to examine instead of a vague idea about “AI for agencies.”

Ras added a useful habit here: prototype to understand, not just to ship. Before using a tool such as Daytona, a sandbox provider, he might spend 15 minutes trying to build a rough version himself. He doesn’t expect to replace the real product. Building the prototype helps him understand the problem.

That’s a different way to use cheap experimentation. The prototype can be valuable even if you throw it away.

3. Find underpriced local marketplace listings

Some opportunities are less about launching a venture-backed startup and more about making useful extra income.

Monitor Facebook Marketplace and Craigslist in my city for cameras, furniture, and bikes listed substantially below comparable prices. Send me an alert with the listing link when a promising opportunity appears.

The agent’s job is to reduce the time spent searching. You still need to judge the item, the seller, the condition, and whether there is actually a resale opportunity.

I like this because it applies AI to an existing activity people already understand. It could be useful for someone looking for extra income, and some people genuinely enjoy finding and reselling things. It’s an idea to test, not a guaranteed daily earnings strategy.

4. Build a one-person-company dashboard

A founder can have plenty of data and very little clarity.

The answers are scattered across notes, Stripe exports, analytics, customer calls, and project lists. Ask the agent to bring those sources together around decisions.

Use my documents, notes, Stripe exports, analytics, customer calls, and project lists to create a weekly operator dashboard. Explain what is making money, what is wasting time, what customers want, and what I should stop doing. End with the three highest-leverage actions for next week, supported by evidence.

The important part isn’t the dashboard’s appearance. It’s whether it helps you decide what to do on Monday morning.

“Be blunt and show your work” is the spirit here. I don’t need another reassuring summary. I need a useful operating picture.

5. Audit the company for agent opportunities

Before I hire someone, I increasingly ask: should an agent do this?

Sometimes the answer is yes. Sometimes it absolutely isn’t. The point is to make that a deliberate decision.

Review how this business operates and identify tasks we could give to agents before adding another hire. For each task, estimate human time, the cost of mistakes, required tools, automation difficulty, and the first safe version we could deploy. Prioritize opportunities that could save money or create revenue within 30 days.

This is better than asking, “What can we automate?” It forces you to consider the consequences of mistakes and the smallest useful deployment.

The best first agent isn’t necessarily the most impressive one. It’s the one with a clear job and a manageable failure cost.

6. Use the browser to discover the automation

Computer use is one of the capabilities Ras was most excited about. He described using Astra through Codex to navigate the Canada Revenue Agency website and find specific information he hadn’t been able to locate through Google.

That matters because plenty of business work lives inside awkward websites, not clean APIs.

Complete this workflow in the browser. Navigate the actual sites, collect the required information, and fill forms where appropriate. Record anything that breaks or slows you down. Return the completed output, a repeatable standard operating procedure, and a plan for turning the workflow into an agent.

You get three things from one exercise:

• The immediate task completed.
• A documented process someone else can follow.
• A clearer specification for future automation.

Instead of designing an automation around how you imagine the process works, you learn from doing the actual work.

7. Create a recurring QA workflow

Quality assurance shouldn’t only ask whether a button works. It should also ask whether a person understands what happens next.

Every night, test my app on a real phone. Go through signup, checkout, and the main user flows. Capture screenshots of anything broken or confusing, and explain the problem.

“Confusing” is doing a lot of work in that prompt.

A technically functional screen can still have too much information, a poorly placed offer, or an unclear next step. Those are product problems too.

The intended workflow includes real-device testing, which means you need the appropriate device access and setup. The useful idea is recurring testing that looks for both functional bugs and UX friction.

8. Track competitors without obsessing over them

There are founders who spend too much time thinking about competitors. There are founders who ignore them completely.

A monthly report is a pretty good middle ground.

Review my top three competitors through their products and customer emails. Send a monthly report covering new features, pricing changes, and areas where their experience is better than ours.

You don’t need a daily emotional relationship with every competitor’s release notes. You need to know what changed, what matters, and whether there is anything worth learning.

Then get back to building. Innovation still matters more than imitation.

9. Build a game with a business purpose

Yes, there is a game prompt on this list. But the goal isn’t to prove the model can make a game.

The goal is to create something people want to share that also connects to your product.

Build a browser game around this mechanic. Include progression, tension, scoring, failure, and polish. Give people a compelling reason to send it to a friend. Add email or SMS lead capture that connects naturally to my core product.

That reason to share is critical. A game can be technically impressive and still give someone no reason to send it anywhere.

The question is not, “Can we build a first-person shooter?” It’s, “Can we make something fun that introduces the right people to what we sell?”

From an AI Speaker Idea to Parts, a Layout, and Code

The hardware example is where this gets really interesting.

Ras has wanted a physical home for his personal AI agent for a long time. Think less “chat window” and more “Jarvis in the room.”

He likes the HomePod’s industrial design but wanted his own agent behind the experience. He first investigated whether he could modify a HomePod. When that didn’t look like a workable route for him, he asked a bigger question: what if he built his own speaker?

Step 1: Make the idea visible

He had Blender installed, so he asked the agent to establish a connection and create a HomePod-like speaker model.

Then he supplied the repository for his personal agent, Ruth, and asked how to build a DIY AI home speaker around it.

The agent proposed a Raspberry Pi-based setup and explained that he could load his own agent onto it. Suddenly, an intimidating hardware idea had a possible implementation.

Step 2: Turn the concept into a shopping list

The agent produced an initial component plan and a budget of roughly $350 to $450 before tax. Ras lives in Canada, so he asked for Amazon.ca links to make the shopping list locally useful.

The resulting order came to about CA$561 before tax, and he bought the parts.

That amount was a prototype parts order, not the cost of manufacturing a finished consumer product. But it was enough to move him from imagining the project to committing to a first build.

Step 3: Ask for an assembly visualization

Buying the parts created the next obvious problem: he didn’t know how to put them together.

Rather than stop, he asked the agent to use Blender to show the project’s setup. It laid out the board, audio components, and connections in a spatial arrangement he could understand.

This is an underrated use of 3D generation. The model isn’t just making something attractive. It’s helping someone understand a physical system.

For background on the computing platform, the Raspberry Pi documentation is a useful companion reference. The agent can guide the project, while the documentation gives you a place to check platform-specific details.

Step 4: Prepare the software

Ras then asked what needed to change in his existing agent repository to run the speaker setup.

The agent wrote code, opened a pull request, and explained the changes. Ras merged it.

It also helped explore the next stage: what a more polished enclosure would require and which suppliers he could contact about making the shell.

The sequence was roughly half an hour of work:

1. Explore the speaker idea.
2. Create a visual concept.
3. Identify a plausible hardware setup.
4. Source the prototype components.
5. Visualize how the components fit together.
6. Prepare the software changes in a pull request.
7. Outline a path toward a custom enclosure.

The physical assembly and end-to-end validation were still ahead. That’s important. But the project had moved from “I wouldn’t know where to start” to “I have parts coming and a plan to follow.”

Use the Intelligence to Expand What You’ll Attempt

Ras put his finger on something: model intelligence keeps climbing, but people’s willingness to attempt unfamiliar things hasn’t necessarily kept up.

His speaker wasn’t the only example. He also had an unused PS4 and asked what he could do with it besides gaming. The conversation led toward using it as a Linux learning project, potentially something he could access over SSH.

He even explored whether a Linux distribution such as Omarchy could fit into that project. That was exploration, not a demonstrated working installation. The significant shift was that he was asking questions about operating systems and hardware that sit outside his usual web and mobile development work.

That’s the broader opportunity. You can use an agent to help investigate unfamiliar territory rather than staying permanently inside your current skill set.

And the project doesn’t have to become a business to be worthwhile. You can learn something, develop a new skill, or discover an opportunity you wouldn’t have recognized otherwise.

From Vibe Coding to Vibe Manufacturing

When tools such as Lovable, Bolt, Replit, and v0 gained momentum in 2024, the number of people willing to build software expanded.

People who would never have opened a traditional development environment started making websites, apps, and prototypes. The effects of that shift are still playing out.

I think we’re approaching a similar shift for physical products.

Call it vibe manufacturing: using AI to help move from a product idea into models, parts, code, and a prototype plan.

That doesn’t mean manufacturing is solved. A render is not a working product, and a parts list is not a production line. What changes is the intimidation factor and the amount of help available at the beginning.

You can start with a printable object, a CAD file, a Blender model, or a circuit-board concept. Something concrete enough to test the next assumption.

My business thesis is that this could produce thousands of niche companies doing $1 million or more in annual revenue. That’s a prediction, not an established outcome. But the underlying idea is simple: a market doesn’t have to be enormous if you can build a focused product economically.

The questions I’d ask are:

• What small physical product solves a real, specific pain point?
• Can I prototype it without making a huge upfront commitment?
• What do I need to learn before deciding whether it deserves a business?
• If it works, could targeted Meta ads help me reach the right customers?

That’s much more interesting to me than making a generic object because the model can.

Measure Value by the Work You Unlock

Ras closed with an example of the subscription economics. He asked his agent to estimate the API-equivalent cost of his recent usage. It returned a figure above $24,000 for one week, while he was paying for a couple of $200 subscriptions.

That was an agent-generated estimate, not an independently verified invoice. Still, it explains his enthusiasm about the amount of intelligence available through subscriptions at the time.

His personal preference had shifted strongly toward Astra, enough that he said he would choose it over his Claude subscription if forced to pick one. That’s a user’s judgment, not a universal ranking.

The more useful takeaway is what to do with the access you have.

Don’t spend all of it asking for work a much smaller model could handle. Audit the product that already earns revenue. Map the service you might turn into software. Investigate the physical product you assumed required a completely different career.

Pick one idea that felt too large last year, and ask for the first concrete step.