From Six Seven to Six Zeros to What Next?

Besides 6-9 and 6-7, I've noticed that 6-0 has become the internet's favorite number over the past couple of years.

Engineers used to get excited about 6-9, although for a slightly different reason than everyone else.

An engineer would proudly say, "I love six nines."

And a kid will say, "6-7."

Now engineers are bragging about how much they love six zeros, and surprisingly, everyone else agrees with them again. Except this time, they're actually talking about the same thing. Well, sort of.

For anyone confused about what I'm talking about, let me explain.

Why Six Nines Mattered

In software engineering, six nines means 99.9999% uptime. Basically, it's a measure of how often an app or website is working instead of being down.

Imagine your banking app suddenly stops working. You can't transfer money, check your balance, or pay someone. Annoying, right? Now imagine that happening to millions of people at once. That's the kind of nightmare engineers are paid to prevent.

Here's why six nines was such a big deal:

• 99% uptime means your app could be down for almost four days a year.
• 99.9999% uptime means it could only be down for about 31 seconds a year.

Getting from four days to 31 seconds is incredibly difficult. Servers crash, bugs happen, and sometimes someone makes a tiny mistake that brings down an entire service.

That's why engineers used to brag about their nines. Keeping something running almost perfectly was a serious achievement.

The New Obsession: Six Zeros

Fast forward to today, and engineers seem to have found a new favorite number: six zeros. More specifically, the zeros after the decimal point on an AI model's pricing page.

For anyone unfamiliar, AI models charge based on something called tokens. Think of tokens as small pieces of text that an AI reads or generates. The more you use, the more you pay. And since a single token can cost a tiny fraction of a cent, you'll find an impressive number of zeros on those pricing pages.

Here's what some AI models cost per 1,000 tokens, in US dollars (input / output):

• GPT-5 nano — $0.00005 / $0.00040
• Gemini 2.5 Flash-Lite — $0.00010 / $0.00040
• DeepSeek V4.1 Flash 1017 — $0.00004 / $0.00050
• Claude Haiku 4.5 — $0.00100 / $0.00500
• Claude Sonnet 5 — $0.00200 / $0.01000

Illustrative API text prices, converted to per-1,000-token prices. Verify model names, availability, and current rates before publication. Input is text sent to a model; output is text generated. Reference: OpenRouter.

Look at those beautiful zeros.

To be fair, AI companies usually advertise their prices per million tokens, which makes a lot more sense than whatever I've done to this table.

But I think you get the idea.

My 13-hour screen time has exposed me to people showing off their Hermes workflows, explaining how ridiculously cheap they are to run with their favorite models. Another person shares a detailed comparison of seventeen models, complete with benchmarks, pricing, and a recommendation for which one gives you the most bang for your buck.

And of course, some comments are saying, "Why use that model? This one's cheaper." Or, "You could've done the same thing for half the price."

Or there's someone promoting a free model you've never heard of, conveniently followed by a referral link promising extra credits.

I find it funny that we've reached a point where people are debating whether spending another fifty cents on AI was financially irresponsible.

I'm not saying cost doesn't matter. It obviously does, especially when you're running thousands of requests. But the amount of attention we give to those tiny numbers is fascinating.

And I think there's a reason why so many people suddenly care about them.

Building Is So Cheap, Let's Do It!

Since AI is getting so cheap and accessible, more people are trying to build software and agents on their own.

Someone has built their own ERP system over the weekend. Someone else automated their entire workflow with AI. Another person made a personal expense tracker because none of the existing apps worked quite the way they wanted.

Then you see some guru claiming to run over 100 AI agents. I still can't imagine what all those agents are doing.

What's interesting is that not all of these people are software engineers. Some are just curious people, business owners, students, people with domain expertise, or people who happen to have too much time on their hands.

Just a few years ago, building an application or automating a business process could be a serious investment. You needed technical knowledge, developers, time, and infrastructure. Even automating a repetitive task for a small business could mean hiring someone or paying for an expensive subscription.

If you had an idea but couldn't code, your options were pretty limited. You either paid someone to build it or spent months figuring out how to do it yourself.

Now you can open an AI assistant, describe what you want in plain human language, and watch it put something together.

And apparently, building something from scratch isn't the only hot topic. Recently, REA (Reverse Engineer Anything) also went viral. It's a tool that helps AI coding agents inspect existing software and figure out how its features work, even without having the original source code.

People have been experimenting with reconstructing applications and features from commercial software, including the kinds of tools made by Adobo and Marcohard. We're moving beyond asking AI to build a new app. Now people are asking it to figure out how an existing app works so they can make their own version.

Imagine spending years building a SaaS product, only to see someone on the internet announce that they recreated one of its main features over the weekend.

And of course, the comment section probably wants to know which model they used and whether a cheaper one could've done it.

I find this development both fascinating and a little terrifying for anyone trying to make money selling software.

To be fair, reverse-engineering a feature doesn't mean you've recreated an entire product or business. There's still infrastructure, customer support, security, reliability, and all the other things people only notice when something breaks. Not to mention intellectual property and licensing questions.

And I would trust my vibe coded app with my savings because it's mine and im proud (big grind)

But I still find it remarkable how quickly we've gone from needing to learn the language of computers to simply telling computers what we want.

We've reached a point where people can practically summon software into existence with a spell (re: prompt).

And now we're even teaching those spells to study someone else's software first.

Agent Boom: I Want to Be Forever Young Cheap

With AI becoming so affordable, people are building their own applications, replacing existing tools, and automating parts of their work that previously required paid software.

And I can understand the excitement. Why pay for another subscription when you can build something that does exactly what you need?

But there's something funny about this new independence.

We're building our own software to avoid depending on someone else's software, only to build it on top of AI models and services whose prices, usage limits, and availability we don't control.

We might have escaped one subscription just to end up with three different API bills. Oh wait, there's another service for that too. XXRouter or XXAI lets you access multiple AI models through one API. (Please endorse me.)

Great. We now have a service to manage the services we use to replace other services.

And that's what makes me wonder how long this cheap AI party will last.

There's no free lunch, except MBG perhaps (an Indonesian joke). That's not even actually free!?

AI providers are constantly experimenting with pricing, usage quotas, and subscription plans. Even while writing this article, I've already had to revisit the pricing table. The numbers aren't something we can assume will stay the same forever.

I've also been paying attention to changes in how AI coding tools and assistants allocate usage. I mainly use ChatGPT for brainstorming and Codex for coding, so any changes to how those limits are shared could affect my workflow.

Meanwhile, other services are experimenting with alternative billing models. Some charge by tokens, others by subscriptions, and some explore usage measured in time.

It makes me wonder whether we'll eventually move toward running more AI models locally, especially if people begin relying on their personal agents for everyday work.

Although, looking at GPU prices, I'm not sure self-hosting would solve my financial problems either.

So here's the part I keep thinking about: are we actually becoming more independent by building our own software, or are we just trading one dependency for another?

What Are We Chasing Next?

We used to chase six nines because building reliable software was a serious engineering achievement.

Now we're chasing six zeros because AI has made creating software accessible to more people than ever.

But the more things we build, the more questions I have.

• What's the right price for AI models, and will they stay affordable?
• Will self-hosting become a reasonable alternative to cloud models?
• Will hardware become cheaper, or will my dream PC remain a dream?
• What comes after the obsession with building personal AI agents?
• If everyone can build their own tools, and potentially recreate features from software they already use, is it still worth building another SaaS?

And maybe the question I'm most curious about is whether all this cheap software is actually making us more independent.

I don't really know.

For now, my answers are: IDK, I hope so because I want a new PC, I'm going to ask my AI, and yes, probably.

But perhaps I should ask one more thing.

When everyone can build software and even hardware on their own, what are we chasing next my fellow SWE? Ternak lele kah?