The AI Buffet Closed. Now the Meter's Running.

Put a hard ceiling on your AI spend before you ever hit "go": a monthly cap, a per-run budget, and the right-sized model for the job. Do that and a runaway agent cannot eat your month in an afternoon.

That is the whole defense. The more interesting question is why you suddenly need one. Flat-rate AI quietly died this summer, and most people are still working like it is all-you-can-eat.

The deal changed while nobody was looking

For about two years, the arrangement was simple: pay twenty bucks a month, or two hundred if you were serious, and use the thing as much as you wanted. The meter did not matter because you could not really move it. You typed a question, you got an answer.

Agents broke that. An agent is the AI running in a loop, taking actions on its own instead of waiting for your next message. Once it runs in a loop, it stops sipping and starts gulping. One research firm actually measured it: they bought every subscription tier, ran agent workloads flat out, and mapped the usage to real token prices. An agentic workload can burn up to a thousand times more tokens than a chat exchange, and a two-hundred-dollar plan pushed to its limits racked up around fourteen thousand dollars in underlying compute.

You will not hit fourteen grand. But you can absolutely hit a number that ruins your Tuesday, and the tools are now built to let you.

Tokens are the meter, and agents floor it

A token is the unit AI gets billed by: a chunk of text, roughly a word or a bit less. Every word you send and every word the model sends back gets counted and priced. In chat that is a few hundred tokens, trivial. The part that surprises people: an agent re-sends the entire conversation on every turn. Every file it opened, every failed attempt, all of its own verbose output, mailed back to the model at each step so it remembers what it is doing. Picture a snowball rolling downhill with labels sticking to it: file dump, error log, retry, retry again. A session that started with three sentences of typing can quietly balloon into millions of input tokens. You did not type them. The loop did, while you got coffee.

This is not a prediction; there are dates on it

On June 1st, GitHub Copilot moved every plan to usage-based billing: plain autocomplete stays included, but agent work draws down metered credits. Anthropic went further, then flinched: it announced that programmatic usage, meaning agents and scripts calling the model automatically, would move to a separate full-price meter on June 15th, and then paused the change the day it was supposed to go live. The exact price of running an agent got announced and reversed inside a month. That volatility is the point. In one survey of nearly three thousand developers this year, unpredictable cost overtook security as the number one pain point.

The line that decides your bill

Vendors now distinguish interactive usage (you, typing and reading, human-paced, still subsidized) from programmatic usage (the machine using the machine, metered, because a loop never gets tired). Whatever the logos do next quarter, that line is not going away. Knowing which side your work sits on is how you predict your bill instead of getting surprised by it.

Five controls, in order of impact

1. Right-size the model. The biggest model can cost five times (or more) per token what the small one does. Reformatting, simple edits, and first drafts go to the cheap model; save the expensive one for genuinely hard reasoning.
2. Set a hard cap and a per-run limit. A monthly ceiling plus a per-run budget flag on automated jobs; the per-run limit is what saves you when a script loops at 2 a.m. A cap you have to remember to check is not a cap.
3. Keep sessions short. Context re-sends every turn, so a long messy session is a bill multiplier. Finish the task, start fresh. Stable prompts also get cached by the vendor, which can cut repeated input cost by around ninety percent. Ideally, keep session context beneath 160k, but up to 200k is still fine for costs.
4. Know your interactive-versus-programmatic mix. If most of your spend is automated jobs, you are on the expensive side of the line; watch it closely.
5. Read the usage dashboard weekly. Five minutes on the calendar. The bill that blows up is always the one nobody looked at until the statement arrived.

Here is the uncomfortable follow-up question: if an agent can quietly multiply your bill, what else in your setup is it quietly multiplying? Cost is just the most visible symptom of a codebase that is not ready for agents. The readiness quiz scores the rest of it in about five minutes, no email required: agentreadyscore.ai.

Watch the meter. It is watching you.