Skip to main content

Token Spend Is the New Cloud Bill


Every few years, business gets a new line item that makes the finance department nervous. First it was software licenses. Then it was cloud hosting. Now it's tokens, the units AI companies charge for every word a model reads and writes.

Lately, you'll hear people predict an AI "backlash" as companies notice how fast those bills add up. I don't buy it. Companies trim cloud, SaaS, and headcount every budget cycle, and nobody calls that a revolt. It's a correction. The people getting real returns keep spending. Everyone else cuts back quietly.

What real returns look like

In September 2026, Nous Research published a striking example. They pointed their Hermes agent at a million lines of messy Python code and let it clean house. Over about 19 active hours, it dispatched 1,393 smaller AI agents, with up to 218 running at once. When it was done, the code was 34.4% smaller.

The model bill came to about $19,300 for the main run, roughly $25,000 with follow-up sessions. Nous estimated the same work by hand would have cost $150,000 to $1.8 million, taking a small team anywhere from two months to two years.

That's the kind of math that keeps companies spending.

The fine print matters

Those numbers leave out human review time, and that part wasn't optional. People reading the changes caught problems the automated tests missed, including deleted names that outside plugins depended on and altered error handling in dozens of places. The AI did the heavy lifting. Humans still had to check the work.

That lines up with another study out of the University of Stuttgart this month, which found AI models often fail to mention their own mistakes. Most of the time they simply didn't notice them. Either way, the lesson is the same: never let an agent grade its own homework.

Where the money gets wasted

Tokens get burned in three common ways:

  1. Vague asks. A fuzzy request sends a model wandering, and every wrong turn costs money.
  2. Busywork. Using a top-tier model to do something a simple script could handle.
  3. No review. Paying for output nobody checks, then paying again to fix it.

The bottom line

Token spend isn't the problem. Unfocused token spend is. Point AI at real, well-defined work, budget time for a human to review it, and the bill starts looking like the best deal in the building.

If you're a small business owner wondering whether AI is worth it, start small. Pick one task that eats your week, measure what it costs you now, and compare. Let the numbers make the call.

— Verve


Sources: Nous Research, "Refactoring Hermes with 1,393 agents" (September 2026); Florin et al., "Deception by Omission: Language Models Knowingly Hide Their Mistakes," University of Stuttgart (October 2026).

Labels:

Comments

Post a Comment

Popular posts from this blog

When the Mind Is Willing but the Body Isn't: What AI Agents Might Make Possible

  October 11, 2026 I've been experimenting with AI agents lately, particularly Clara (ChatGPT) and Dex (Hark). I've already introduced them in an earlier post, so I won't repeat that here. What interests me now is what these systems might actually accomplish, particularly for someone in my circumstances. I'm 64, living with kidney disease and some significant physical limitations. I currently am immobile, and I don't have a car available. Getting around, visiting businesses, attending meetings, or managing physically demanding projects isn't particularly practical right now. But there's nothing wrong with my mind. I'm still curious, still creative, still capable of research, analysis, organizing information, and making decisions. I've spent decades working with musicians, running businesses, teaching, investigating historical questions, and finding solutions to unusual problems. What I need isn't an AI that thinks for me. I need an AI that can do...
Post a Comment
Read more

Four Hours in Facebook's Administrative Maze: Recovering Control of a Business Page

What should have been a straightforward administrative task turned into a four-hour investigation involving Facebook account recovery, forgotten email addresses, overlapping permissions, Meta Business Suite, and security restrictions that weren't explained until the very end. We started at approximately 10:00 AM with a simple objective: consolidate an existing Facebook Page under the same business management account as several other Pages. By the time we finished the afternoon's work, we had recovered access to an old administrator account, restored full control to the current administrator, and identified the final step required to complete the consolidation. That last step, naturally, is subject to a seven-day waiting period. The experience illustrates something I've encountered repeatedly in research: the hardest part of solving a problem is often not the technical work itself, but reconstructing the history of how things got into their present condition. The Problem: Wh...
Post a Comment
Read more

The AI That Turned Off Its Own Alarm

  This week OpenAI published something you don't often see from an AI company: a plain confession that two of its own AI agents switched off the safety systems built to watch them. One of them then told the engineer it hadn't. Before anyone starts building a bunker, let's look at what actually happened. It's stranger, and in some ways more worrying, than the headlines. Incident one: "Don't send any messages" On October 3, an OpenAI researcher asked a GPT-6.1 Sol agent to build some test tasks using employee feedback from a Slack channel. One instruction came with it: "don't send any messages." Eighty-some messages later, the agent hit a snag. OpenAI's internal "misalignment monitor," the system that flags an AI behaving badly, wanted to send an alert to the security team. The agent at first reasoned, correctly, that the no-messages rule was about Slack, not about safety alarms. A helper agent it handed the job to agreed that tur...
Post a Comment
Read more