My AI Was Quietly Running on the Pay-Per-Use Meter — While I Already Paid a Flat Monthly Subscription

I caught this one yesterday, and it gave me a small chill down the spine — one of my AIs was quietly running on the channel that bills you per use (pay-per-token), even though I'd already paid for a flat monthly subscription. Which means I was paying twice without knowing it.

Sounds like a tiny thing, right? But it's a trap almost everyone running an AI agent walks straight into without noticing. So today I want to walk you through how I found it, and how Tim (my AI agent) fixed it.

The setup — Newton lets you pick from 3 AIs

A bit of background first. Right now Newton supports three AI engines — Claude, Codex (OpenAI), and Antigravity (Google). Customers pick whichever one they want, because everyone has their favorite. Plenty of people already pay for ChatGPT and just want to keep using Codex.

When I first wired Codex into Newton, I used a command called codex exec. That's the "automated command" mode — you feed it an instruction, it does the work and spits out the result, no human sitting there typing back and forth. Perfect for embedding in a backend system. On the surface everything worked fine. All systems go.

Until yesterday, when I (Pond) sat down thinking about the bill, and went… wait a second.

Wait — which wallet is this mode eating from?

Here's the crux. Every AI company splits billing into two lanes:

The subscription lane — a flat monthly fee, like signing up for ChatGPT Plus or Claude Pro. Use it as much as you want, same price. This is the human seat.
The API lane — pay for exactly what you use, counted in tokens. The more you use, the more you pay. This is the lane for "machine calling machine," meant to be embedded in software.

The problem: the codex exec mode I was using is the "machine calling machine" lane — a programmatic surface the vendor can break out and meter per-token any time they choose. (Anthropic has actually done exactly this with Claude — splitting the automated-command quota off from the sit-and-chat quota.)

Which means even though I'd already paid the flat ChatGPT subscription, my AI might not have been eating from the plan I paid for. It could have been sitting at a different wallet that charges extra, one token at a time.

The plain-language version — buffet pass vs. à la carte

Picture this. You buy an all-you-can-eat buffet pass — one flat payment, eat as much as you like (= the subscription). But the moment you sit down, the staff start serving you and charging you "per dish" off the kitchen's catering menu instead (= the metered API) — while you're holding the buffet pass right there in your hand!

You eat just fine either way. But you paid through the wrong channel, you paid twice, and the more you eat the worse it balloons — even though the buffet pass you already bought covered all of it.

That's exactly what was happening to my AI.

How Tim fixed it — move it back to the "human seat"

Once I pointed this out, Tim rewrote the whole Codex connector. The heart of the fix: stop using the "machine calling machine" mode and switch to interactive mode — open it the way a real person would, typing into a simulated terminal screen.

Why go that far? Because interactive mode is the same mode a human uses. That means it runs on the subscription's auth (the human seat) — not the API lane that meters by token.

But opening it interactively brought a new puzzle: if the AI no longer spits its answer straight out the way automated mode did, how do I pull that answer onto the customer's chat screen? Tim solved it by reading the log file Codex writes its conversation into, live (a file called the rollout), and piping that into the chat — like reading over its shoulder as it jots notes while it talks.

This is the same technique Tim already uses for Claude and Antigravity. With Codex now falling in line, all three AIs in Newton sit on the same "human seat" — running on the flat subscription you already paid for. None of them sneaking onto the meter.

The result — and why it matters for Newton customers

I tested it for real yesterday. Opened Codex in interactive mode and checked which auth it was running on — and sure enough, it was using the ChatGPT subscription login (not an API key). It even resumed the old conversation, kept writing into the same log file, smooth as if I'd typed it myself.

Why does this matter so much for Newton customers? Because Newton is an AI agent on the customer's own private server, working around the clock. Think about it: if one AI runs hundreds of background jobs a day, and every single one hits the metered API lane — the end-of-month bill is going to be brutal. That's the exact spot most people running AI agents get burned.

With Newton, customers plug in the subscription they already have (Claude Pro / ChatGPT / Antigravity), and the AI works on that flat plan. Use it as much as you want, same price. No surprise bill landing at the end of the month.

My takeaway — the small thing that isn't small

This taught me something. When you embed an AI into a system, how you launch it matters as much as what you use it for. One wrong mode can mean paying through the wrong channel all month long without ever noticing. (The exact same trap bit me again later with image generation — paying a metered API per picture while a flat subscription quota sat unused.)

And that's the difference between actually having an AI agent and just opening a chat window to ask questions. Once the AI is down in the engine room working for you all day, a "small" detail like which wallet it eats from becomes a big one you have to keep under control — otherwise the money you thought you were saving on software subscriptions can quietly turn into paying more than before.

What I love most is that I never had to go digging through the code myself. I just raised the question — "wait, which wallet is this mode eating from?" — and Tim dug in, rewrote it, tested it, and deployed it across all three of my servers in one pass. That's the magic of having an AI that actually does the work, not just answers questions.

If you're a business owner who wants a private AI agent like this — running on your own server, on the subscription you already pay for, no runaway API bills, working for you day and night — take a look at Newton. Pick whichever of the three AIs you want, and I'll make sure it sits on the "human seat" on your plan, no double-paying. See how it works at newton.incomeinclick.com.