I Don't Think the Token Subsidy Is Going Away. And That Says a Lot About the Future of AI

There’s an idea I’m hearing more and more: that this era of “subsidized tokens” is going to end.

That right now OpenAI and Anthropic are letting us use ChatGPT, Codex, or Claude at ridiculously low prices compared to the API, but that’s just a temporary phase. That once we’re all hooked, we’ll have to pay the “real price.”

I don’t see it that way.

I’m not saying they won’t change limits, plans, or conditions. Of course they will. I’m also not saying they’re going to give away infinite intelligence forever. That would be naive.

What I’m saying is something else: I don’t think the subsidy will disappear. I think it’s going to become a structural competitive advantage.

Intelligence Gets Cheaper, Even If the Product Doesn’t

My sense is that a lot of people are mixing up two different things:

The real cost of producing intelligence
The price of the product we buy

And they’re not the same.

Intelligence, understood as inference capability, is on a very clear trend: getting cheaper over time. Today’s frontier models are still expensive, yes. But their immediately lower-tier models are dirt cheap compared to what they cost a year or two ago. And that’s the same pattern we’ve seen over and over in technology: what is elite today becomes commoditized tomorrow.

That’s why I don’t think the likely scenario is: “the party’s over, now everyone has to pay API-level token prices.”

I think the likely scenario is this:

the internal cost of serving those models will keep dropping
companies will try to get us to use them for more and more things
and the visible price of premium subscriptions won’t change much

In other words: they’re not going to kill the subsidy; they’re finally going to make it work for them financially.

And that’s very different.

The Business Isn’t Selling You Tokens. It’s Selling You a Habit

OpenAI and Anthropic aren’t just building models. They’re building products.

And a product isn’t optimized so you understand the marginal cost of each request. It’s optimized so that:

it’s useful to you every day
you incorporate it into your work
you depend on it more and more
and moving up to the next plan makes sense to you

That’s why I think the natural move isn’t to start charging us the API price “as is,” but to try to move us up to higher and higher plans.

Now there are tiers like 20 and 200. My impression is that the goal isn’t to destroy that scheme, but to make it seem reasonable for more people to live on the higher tier. Especially programmers, technical teams, and heavy users.

That fits much better with how software companies work.

They’re not selling you cost units. They’re selling you a value proposition you want to keep paying for.

Plus, They Can’t Get Too Cocky

Here’s another important point.

Even if they wanted to crank up the price, they’re not alone in the market.

OpenAI and Anthropic are in a fierce war. Google is still around. And the Chinese models are coming on strong, with a quality/price ratio that makes it very hard to get too arrogant.

Add to that another, slower but real pressure: local models are getting better. They’re not ready to replace everything in most workflows yet, but they do set a competitive floor. And, above all, they force the idea of a “reasonable price” to stay grounded in reality.

Today, a lot of people don’t seriously consider switching to cheaper models because the top models are still worth it and the cost is still manageable.

But if the price really shot up, that would change fast.

People would start to:

try out Chinese alternatives
optimize workflows to use less
use cheaper models for more steps
and increasingly consider running things locally where it makes sense

That’s why I don’t buy the argument that “once we’re hooked, we’ll pay whatever it takes.”

No. We’ll pay a lot if it’s still worth it. As soon as it stops being worth it, we’ll look for something else.

And they know it.

So Is AI a Bubble?

My opinion: not in the simplistic sense a lot of people mean.

Is there hype? Of course. Is there a ton of smoke and mirrors? Also yes. Are there companies cramming “AI” into every slide just to raise money? No doubt about it.

But a real bubble, in the strong sense, is usually built on a huge promise with very poor practical utility.

And in programming, honestly, that’s not what we’re seeing anymore.

Since November, we’ve seen a very serious shift.

I’m not talking about benchmarks. I couldn’t care less.

I’m talking about real, day-to-day use. What happens when you sit down to work.

A few months ago, the model fit pretty well as a copilot. Useful, yes. Sometimes brilliant. But still very dependent, very fragile, very prone to needing constant micromanagement.

Now we’re entering a new phase.

I’m not saying it codes on its own or that we can disappear from the process. But we are starting to accept that we can give it a lot more leeway, let it tackle bigger blocks, orchestrate it, supervise it, and correct it instead of dictating every keystroke.

That’s not a copilot anymore. That’s starting to look like a delegate.

And that’s no small change. It’s a regime change.

Thinking We’ve “Hit the Ceiling” Doesn’t Match What We’re Seeing

Lately, I also hear a lot that LLMs have hit their ceiling. That Transformers have given all they can. That the big stuff has already happened.

I don’t see any practical evidence of that. At least not in programming.

If we were really at a functional ceiling, we wouldn’t have seen such a clear leap in real utility in just a few months.

And here’s an important nuance: part of the leap surely comes from better models, yes. But another part comes from something just as important, if not more so:

better training for development tasks
better tools
better contexts
better harnesses
better ways to maintain state
better workflows around the model

So: even if someone wanted to argue that the “pure model” isn’t improving that much, they’d still have a problem.

Because the complete system is improving.

And in the end, what we use isn’t the model in isolation. What we use is the whole system.

Even If Models Froze Today, We’d Still Have Years of Improvement Ahead

For me, this is the most important point of all.

Imagine, for a moment, that tomorrow progress on the base model stopped. No better GPT, no better Claude, no better Gemini. We’re left with exactly what we have today.

Would that stop real progress in applied AI at work?

I don’t think so. Not even close.

I think we’re still a long way from getting the most out of the current models.

And that margin isn’t just in the model’s “intelligence.” It’s in everything else:

how we give it context
how we break down tasks
how we validate the result
how we do retries
how we integrate tools
how we design interfaces
how we build memory and state
how we organize human work around the agent

Confusing the model’s limit with the system’s limit is one of the biggest mistakes of this stage.

Because maybe a model improves by 15%. But a well-designed system around that model can multiply the real value you get by much more.

That’s why I think we’re still in a very early phase.

Not because “everything is magically going to get better” and that’s it. But because we’re still learning how to really build around these capabilities.

My Bet

My bet, today, is this:

The token subsidy isn’t going away
It’s going to become part of the product
AI isn’t close to hitting its ceiling in programming
We’re nowhere near the practical limit of what we can do with LLMs

What I do think is that we’re leaving behind a very naive first phase.

The phase of “look how impressive this demo is.”

And we’re entering a much more interesting one:

the phase where the question is no longer whether this is useful, but how to design it well to get the most out of it.

And honestly, we’ve got years of progress ahead of us there.