The Strange Golden Age of AI

I grew up when America got online.

I had AOL, and yes, I know AOL was not “the Internet.” There was Netscape, CompuServe, local ISPs, message boards, IRC, and a much bigger world beyond the little welcome screen. But that modem sound is still the sound I associate with the moment ordinary people started getting access to something enormous.

It was messy. Slow. Frequently ridiculous. Nobody really knew what it was going to become.

And that was part of the magic.

Artificial intelligence feels like that to me now. Not because every AI product is good, and certainly not because every prediction about AI is going to come true. There is already plenty of nonsense attached to this technology. Models make mistakes. Companies overstate things. People paste “AI” onto products that did not need it. The social, economic, security, and ethical questions are real.

But step back from the marketing for a moment and look at what an ordinary person can actually get their hands on in September 2026.

It is remarkable.

For roughly the price of a streaming subscription, you can have regular access to some of the most capable general-purpose models ever built. You can give them documents, images, spreadsheets, source code, web research, and long-running problems. Depending on the product, they can speak with you, search for you, write software, analyze files, generate media, work across connected applications, and increasingly take actions instead of merely suggesting them.

And the strange part is that twenty dollars a month is no longer even the interesting part.

The interesting part is how quickly “chatting with AI” is turning into “working with AI.”

A quick translation, if you have not been living in this stuff

The terminology has gotten ahead of normal people.

A frontier model is simply one of the most capable general-purpose AI models available at the time. Think of the top systems from companies such as OpenAI, Anthropic, and Google. A chatbot is the interface most people met first: you type something, it replies.

An agent is different. It is a model given tools and some freedom to work through a task. Instead of only telling you how to reorganize a spreadsheet, it may be able to open the files, inspect the data, write the formulas, check its work, and hand the result back. A coding agent can inspect a software repository, change files, run tests, find that it broke something, fix it, and continue.

An open-weight model is one whose trained model weights are available for you to download and run yourself. You are not necessarily getting every detail of how the model was created, which is why “open-weight” is more precise than “open source,” but you gain a degree of control that a hosted chatbot cannot give you.

Those distinctions matter because they explain why this moment feels different.

Twenty dollars is still an absurd amount of capability

I do not mean “absurd” as in free. Twenty dollars matters. A hundred dollars matters. Two hundred dollars absolutely matters.

I mean absurd relative to what you are buying.

ChatGPT Plus remains a roughly $20 consumer subscription with access to advanced models and expanded use of tools such as Codex and ChatGPT Work. Claude Pro is $20 month-to-month and includes Claude Code along with task handoff, projects, and other tools. Google’s consumer AI Pro plan sits at $19.99 a month and folds Gemini into a broader bundle.

You do not need all three. If you are new to this, I would actively recommend against collecting AI subscriptions before you know what you are going to do with them.

Pick one.

Now think about what twenty dollars historically bought you in software: a productivity app, a streaming service, maybe a small cloud server or a specialized professional tool.

Today it can buy access to a system that moves between writing, research, programming, analysis, tutoring, brainstorming, images, files, and software tools. It can help you begin in fields where the hardest part used to be not knowing enough to ask the first useful question.

It is not automatically right. These systems can be confidently wrong.

But “sometimes wrong” and “not useful” are very different statements.

At $100 or $200 a month, the economics get stranger. Those are not casual subscriptions for everyone, but the high-usage plans offered by the major labs are built for people who work with these systems throughout the day. Anthropic, for example, offers Max plans with 5x and 20x the usage of Pro. OpenAI has similarly moved into higher-capacity Pro tiers.

That is serious money for an individual. It is also surprisingly little money if AI is meaningfully participating in hours of your work every week.

Both things can be true.

The models are improving. The price of intelligence is also falling.

This is one of the reasons I think the current period is easy to underestimate.

Yesterday, September 22, OpenAI released GPT-6 Sol and Luna and lowered API prices substantially for those families. The same day, Anthropic released Claude Opus 5.5 and said it costs 40 percent less to run than Opus 5.

That is not a prediction. That happened yesterday.

The race is no longer simply “who can build the smartest model?”

It is also: who can make intelligence faster, cheaper, easier to use, easier to put inside software, and plentiful enough that people stop rationing it?

When something becomes cheap enough that you stop thinking carefully about every individual use, behavior changes. We saw that with storage. We saw it with bandwidth. We saw it with compute.

We are beginning to see it with intelligence.

Agents are the part I think people are still missing

If your experience with AI is opening a website, asking a question, and reading an answer, I understand why some of the excitement seems excessive.

That experience is useful, but it still feels a little like search with a very sophisticated conversational layer.

Agents are where the shape of the thing changes.

Suppose I want to create a small website.

With a chatbot, I can ask how to build it. I can copy code back and forth. I can ask why something is broken.

With a coding agent, I can hand it the repository and say: understand how this site works, add this section, keep the existing design language, test it, and do not disturb anything unrelated.

Then it can actually work.

That does not mean I stop paying attention. The agent can misunderstand me, make a poor architectural decision, introduce a bug, or spend far too long going down the wrong path. The more consequential the work, the more supervision matters.

But the unit of interaction has changed.

I am no longer only asking for an answer. I am delegating a task.

OpenAI’s Work and Codex experiences, Anthropic’s Claude Code and its newer task-handoff workflows, and the growing ecosystem of agent frameworks are all moving in this direction. OpenAI also launched a managed Agents API this month for developers building long-running, tool-using agents into their own products.

Agents still fail in weird ways. They can wander. They can overcomplicate things.

They also work often enough that this is no longer just a demo.

Then there is the Wild West

The other unusual thing about this moment is that the frontier is not entirely locked inside a few websites.

There are open-weight models from OpenAI, Meta, Mistral, and others. OpenAI’s gpt-oss models, for example, are available under an Apache 2.0 license, can be run on infrastructure you control, and can be adapted or fine-tuned. The smaller gpt-oss model was designed to run with about 16 GB of memory.

That means a technically curious person can run useful AI without asking a hosted provider for permission every time the model is used.

You can choose where the data lives. You can decide what tools the model gets. You can build your own memory system. You can create safety policies and guardrails for your environment rather than accepting one universal set designed for millions of unrelated users.

And that last point deserves some maturity.

Guardrails are not inherently bad. Some are necessary. A model with access to your email, your source code, your bank account, or your production servers should absolutely have boundaries.

The important thing is choice.

For legitimate work, experimentation, research, privacy, or simple curiosity, we now have a spectrum. Use a managed frontier service when that makes sense. Run something locally when that makes sense. Combine the two. Give a cheap model routine work and escalate difficult problems to a frontier model. Put one model in charge of checking another.

A reasonably technical person can do this at home.

That is the Wild West part: not an absence of responsibility, but an unusual amount of freedom before conventions harden.

This is why it feels bigger than the early Internet

The Internet gave ordinary people access to information, publishing, communication, and eventually commerce at a scale that had been difficult to imagine.

AI is building on top of all of that and adding something different: capability.

The Internet could show me a thousand pages about how to write a program. An AI system can help me write the program.

The Internet could give me every public resource about a subject. An AI system can help me synthesize them, challenge my assumptions, identify what I do not understand, and turn the result into something I can use.

An Internet service could expose an API. An agent can increasingly learn how to use that API, combine it with other tools, and execute a workflow.

There is a difference between access to knowledge and access to something that can work with knowledge.

That difference is why this feels so large to me.

I do not think this particular window lasts forever

Calling this a golden age does not mean I think today’s models are the best AI we will ever have. Almost certainly not.

I mean the combination we have right now is unusual.

Frontier models are improving quickly. Consumer access is relatively cheap. Companies are competing aggressively on usage. Agent capabilities are being bundled into subscriptions. APIs keep getting cheaper. Open-weight alternatives are genuinely useful. Standards, business models, interfaces, and regulation have not settled.

Everyone is still trying things.

That creates room.

Maybe competition pushes prices even lower from here. I hope it does. Maybe open models guarantee that a meaningful amount of capability remains outside the control of any one company. I hope that happens too.

But markets mature. Products get segmented. Regulations accumulate. Companies consolidate. Pricing gets optimized. Things currently included because everyone is fighting for adoption may someday become separate products with separate bills.

I do not know which direction every piece goes.

Nobody does.

What I know is that right now the amount of capability available to an individual relative to its cost is astonishing.

We should notice it while it is happening.

If you have not started, do not start by becoming an “AI person”

If you are technical, keep hearing about all of this, and have no idea where to begin, do not spend a weekend comparing twenty models. Do not build a home GPU cluster. Do not learn the vocabulary first. And do not feel obligated to turn your life into a collection of AI workflows.

Get one good subscription.

Pick a real problem you already have and give the model enough context to understand it. Not a trivia question. Something real: a messy spreadsheet, a project you keep putting off, a program you want to write, a set of notes you need to organize, a subject you have wanted to learn, or a recurring process at work that wastes time.

Use it enough to discover where it is strong and where it is unreliable.

Then try an agent on a bounded task. Let it work. Review the result.

That sequence will teach you more than months of AI news.

The useful skill is not “prompt engineering.” It is learning how to collaborate with systems that are capable, fast, tireless, and imperfect.

I want to use this period, not merely watch it

I have more ideas than time. Most people I know do.

Historically, that was simply a constraint of being a person. There are only so many hours. Learning a new field takes time. Building something takes time. Research takes time. Starting from zero is expensive.

AI does not eliminate those constraints.

It changes the exchange rate.

It lets curiosity travel farther before it runs out of hours.

That is what I find so exciting about this moment.

Not replacing people. Not asking a machine to decide what matters. Not turning over judgment.

I mean being able to follow an idea farther than I could yesterday.

I can know very little about something in the morning, ask increasingly better questions, read the underlying sources, challenge the model, prototype an idea, test it, throw it away, start again, and finish the day knowing substantially more than I did when I woke up.

And increasingly, the AI can do some of the mechanical work alongside me while I decide where we are going.

I remember when getting online felt like touching the future.

Eventually the Internet stopped feeling like a destination. It became infrastructure. We stopped announcing that we were “going online” because online became the default state of modern life.

I suspect something similar happens with AI.

Eventually we may stop talking so much about “using AI.” Models and agents will be inside software, operating systems, businesses, scientific tools, creative tools, and things we have not named yet. The technology will become ordinary precisely because it becomes ubiquitous.

But it is not ordinary yet.

We can still see the scaffolding.

We can still choose among radically different approaches. We can still download models, build strange experiments, wire systems together, and discover uses the companies making the models did not anticipate.

We can still be surprised by what twenty dollars buys.

That is why I think this is a golden age.

Not because the technology is perfect. Not because I am certain the future will be better. Not because every AI company deserves our trust or every AI product deserves our attention.

It is a golden age because capability is arriving faster than convention, and because for this brief, strange period, a remarkable amount of that capability is available to almost anyone willing to learn how to use it.

I was lucky enough to watch one technological shift like this happen.

I do not intend to sit this one out.