How I Burned Almost a Billion Tokens in a Month

It's 2026 and we're burning tokens like candles in the mid-1800s. People are quick to blame AI but we're the problem. By our usage alone we are creating the need for more data centers, are we not?

I burned more than half a million in ten minutes. One million in under an hour. And almost a billion in a month !

Detailing my usage

Initially, I started writing this blog around May 10th but due to other projects and commitments I put it on the side. After investigating the root cause the conclusion by the AI model was actually wrong, so I had to dig deeper.

A couple of months ago, in May, Alibaba Cloud offered a 1M-token promotion for Qwen 3.6-plus. I definitely wanted to try it out but I wanted to use it through Codex since I didn't have the subscription at that time.

Not seeing any GitHub repos that already connected other models with Codex, (I don't think any were made at that time.) I wired the Qwen model into Codex through a direct DashScope connection.

That attempt alone burned 23.7M tokens across 19 Hermes sessions and 441 API calls. Then after connecting it to Codex, the first real task consumed almost all of Alibaba's promotion.

I got so upset that almost a million tokens were gone and immediately blamed Alibaba and closed the account! In hindsight, it obviously wasn't their fault for the way I was using it without guardrails etc...

The burn

I gave Codex this task:

"update my portfolio site to match my resume."

Then timeline as follows:

• 21:55 — session opened with 14,297 tokens
• 22:05 — cumulative usage reached 658,057 so about 644K in ten minutes!
• 22:39 — Alibaba email warning fewer than 166,719 tokens remained!
• 22:48 — session reached 1,030,246 tokens!

So, across 30 calls, per-call input grew from 14K to 40–49K. The route recorded 0 cache hits. Of the session's total volume, 1,011,517 tokens 98% were input.

Codex was reading files, editing code, and carrying its growing working context between tool calls.

The root cause

In essence, I drove Codex at full speed with no brakes, and DashScope logged every token.

The route was a blind spot in my GPS Hermes, so to speak, leaving any fuel saving feature useless.

I never hit the emergency brake (my own em dashes, not AI) - the Alibaba guardrail- the quota toggle because I didn't think I needed it.

Almost a billion

The way I use Hermes (and I will post about my setup and why so much), I had processed 989,985,792 tokens across all providers during May, over 27,995 API calls.

From that volume, 81% was cached context, 18% new input, and 1% output.

It wasn't all fresh input or full-price usage. But it was almost a billion tokens processed in one month.

Closing thought

Ideally, yes but we are still at an early stage where it's so easy to misspend tokens, especially free tokens. The governing of Artificial Intelligence is still raw and the infrastructure hasn't been built like other established technologies.

At the end of the day, if I'm to say,

looks like I'm part of the problem that doesn't seem like it's going away