The $20 AI Subsidy Trap (And How to Use the Window)

You pay twenty bucks a month for AI and it feels infinitely cheap. It isn't. The companies behind it are bleeding billions to keep that price on the screen — and they've told us so out loud. This is the same loss-leader playbook Uber ran in 2015, and understanding it changes how you should be using AI right now.

The good news is the part nobody says: this is the cheapest AI will ever be. The subsidy is your window. Below is what's actually happening, and the one free tool that lets you exploit it.

## The free tool that cuts your AI usage ~75%

The single highest-leverage move while AI is still subsidized is to use fewer tokens per task. caveman is a free, open-source token-compression plugin for Claude Code that does exactly that — it compresses the conversation so you get the same work done for roughly a quarter of the tokens.

!BUTTON: Get caveman on GitHub (free)

It's the literal "use AI lean" move this whole article is arguing for. Install it, point it at your Claude Code setup, and your usage drops without changing how you work.

## What's actually happening

AI subscriptions are being sold below cost to lock you in. It looks free-flowing and infinite, but every prompt costs the provider real compute, and right now the provider — not you — is eating most of that cost.

This isn't a theory. The people running these companies have said it on the record, and the financials back it up.

- OpenAI's CEO, on the record (Jan 2025): "we are currently losing money on openai pro subscriptions! people use it much more than we expected."
- OpenAI 2025 financials: roughly $13B revenue against a ~$21B operating loss — about $1.60 spent for every $1 earned.
- GitHub Copilot killed flat-rate pricing (June 2026): moved to token billing; some power-user bills jumped 10x–50x overnight.
- Anthropic added weekly rate limits (Aug 2025), citing "very constrained" compute resources.

| Plan / signal | What you pay | What's really going on |
|---|---|---|
| Consumer AI sub | ~$20 / mo | Heavy users pull multiples of that in real compute |
| Pro / $200 tier | $200 / mo | Verified heavy-user value ≈ 8–14x ($1,600–$2,880) |
| GitHub Copilot | Was flat | Now token-billed; bills up to 50x higher |
| OpenAI (company) | — | ~$1.60 cost per $1 of revenue in 2025 |

## Why give it away? Because someone did this to you before

This is the Uber playbook. In 2015, riders paid roughly 41% of the true cost of a trip — investors covered the other ~59%, burning about $2B a year to build the habit. Cheap rides trained an entire market to stop owning cars and stop calling cabs.

Then the market was captured. Between 2018 and 2021, Uber prices rose about 92%. The subsidy was never charity; it was customer acquisition. MoviePass ran the identical move, burned $200M, and collapsed.

AI is the same trap, sneakier. A single agentic task can fire 10–20 model calls. Your whole workflow quietly leans on it — your code, your writing, your research — and once you can't easily leave, pricing power flips to the provider.

## What most people get wrong

The mistake is assuming the trap is a future per-token price hike. It mostly isn't — inference cost has actually fallen ~40x in two years.

The real lock-in is volume. Unit prices drop, but agentic tools fire so many calls that your total spend climbs anyway, and your dependence climbs with it. So the fix isn't "wait for it to get cheaper" — it's use the cheap window now, and use it lean: fewer tokens per task, tools like caveman, and a workflow you could walk away from if the price ever flips.

## One step to take today

Pick your single most-used AI workflow and cut its token cost this week. Install [caveman](https://github.com/JuliusBrussee/caveman) if you're on Claude Code, or just tighten your prompts and stop dumping whole files into context. The subsidy is real, it's temporary, and the people who use this window deliberately come out ahead of the ones who only notice it when the bill changes.