The Subsidy Window Is Open. Use It.

Right now, you are using AI at a discount you did not earn.

Every time you open ChatGPT, Claude, or Gemini, somebody else is paying the difference between what it costs to serve you and what you pay. OpenAI is projected to burn $14 billion in 2026. Anthropic still runs negative on every dollar of revenue. Sam Altman has openly said the company loses money on the $200 a month Pro plan. The free tier you use on your phone is closer to a billboard than a product.

That window will close. It always does.

You have seen this story before

Uber rides used to cost less than the gas to drive them. DoorDash delivered a $14 burrito for $9. Amazon shipped you a single bottle of shampoo for free and skipped sales tax in most states for a decade. None of that was a gift. It was venture capital buying your habit.

The AI labs are running the same playbook with bigger numbers. They are pricing below cost to capture the market before they have to answer to public shareholders. OpenAI, Anthropic, and xAI are all expected to go public. When that happens, the math changes overnight. Margins have to show up. Subsidies get pulled. Prices climb to whatever the market will bear.

Token costs have already fallen roughly 75 percent in a single year. That is not the labs being generous. That is the labs racing each other to the bottom on your behalf while investors are still writing checks. The next phase looks different.

This is going to be a utility

In five years, AI will sit on your monthly bill next to electricity, water, and internet. Some of it will be bundled into other services. Some of it will be a line item. You will not think about it much. You will just pay it.

The interesting part is what happens in between now and then. Right now, you can experiment with the best models in the world for the price of a streaming service or for nothing at all. You can throw bad prompts at them. You can waste tokens. You can try the same task fifteen different ways until you find one that works. That kind of low stakes practice gets harder when the meter is running for real.

It is exactly like learning to search

When Google showed up, people typed full sentences into it. They asked questions with please and thank you. They had no idea what a good query looked like.

A handful of people figured it out early. They learned how to use quotes. They learned how to chain operators. They learned that a search was a conversation with an index, not a librarian. Those people moved faster than everyone else for the next twenty years. The gap between a good searcher and a bad one was the gap between finding the answer in 30 seconds and giving up after 10 minutes.

AI is in that same window right now. The skill is not memorizing prompt templates. The skill is learning how to talk to a model. How to give it context. How to push back when it goes sideways. How to recognize when it is confidently wrong. How to break a big task into pieces it can actually handle. None of that comes from reading. It comes from reps.

The practice is the point

You cannot learn this from a course. You can take one, and some are useful, but the actual skill builds the way every real skill builds. You use the thing. You get a result you do not like. You figure out why. You try again.

Start small. Use it to plan a trip. Use it to draft a difficult email. Use it to explain a contract clause to you in plain English. Use it to brainstorm a name for something. Use it to debug a recipe. Use it to summarize a long article you do not have time to read. Use it badly at first.

Pay attention to what happens when you give it more context. Pay attention to what happens when you tell it what you do not want. Pay attention to what happens when you ask it to think out loud before it answers. Pay attention to when it makes things up, and learn to feel that coming before it lands.

That feel is the whole skill. It is not transferable from a textbook. It only shows up after you have had a few hundred conversations with these tools and watched what works.

Why now specifically

A few things are true at the same moment, and they will not stay true together for long.

The models are good enough to be genuinely useful for normal tasks. They were not, eighteen months ago. They are now.

The cost to use them is artificially low. That is venture money holding the price down, not the underlying economics.

The interface is still simple. You type. It responds. Once these tools get embedded into every app you use, the surface area gets messier and the skill of talking to a model directly becomes harder to practice.

And nobody is good at this yet. The people who will be considered AI fluent in 2030 are mostly building that fluency right now, in chat windows, on their phones, between other tasks. There is no credential. There is no certification anyone respects. There is just whether you can get useful work out of these tools or not.

What to actually do this week

Pick one task you do regularly that you do not love doing. Writing follow up emails. Planning meals. Reading dense documents. Researching a purchase. Anything.

Take it to a model. Not once. Five times over the next week. Each time, try to be a little more specific about what you want. Each time, push back when the output is generic. Each time, ask it to ask you questions before it answers.

That is the practice. There is no shortcut around it, and there will not be a better moment to start than the one you are in right now, while somebody else is still paying the bill.