> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI Companies Claim $0.35 Per Task. Scientist Measured $210.
- URL: https://wire.fourthweb.ai/ai-companies-claim-0-35-per-task-scientist-measured-210/
- Published: 2026-09-03T16:00:00.000Z
- Updated: 2026-09-03T11:30:48.000Z
- Description: The gap between what AI companies say their tools cost and what they actually cost just got measured—and it's 600x wider than advertised.
- Author: Travis Wright
- Tags: Human Imperative, Agentic Workflows, AI Agents, AI Infrastructure, Compute Wars, OpenAI, Anthropic, Google AI, Microsoft

**The gap between what AI companies say their tools cost and what they actually cost just got measured—and it's 600x wider than advertised.**

### The Summary

- Climate scientist Zeke Hausfather [tracked 1,138 prompts over eight weeks](https://www.fastcompany.com/91600746/how-much-energy-does-agentic-ai-actually-use-one-scientist-tracked-every-prompt-he-sent?partner=rss&utm%5Fsource=rss&utm%5Fmedium=feed&utm%5Fcampaign=rss+fastcompany&utm%5Fcontent=rss) and found his median agentic AI prompt used 150 watt-hours—600x more than the simple chatbot queries [OpenAI](https://wire.fourthweb.ai/tag/openai/) and Google use for their public estimates
- The shift from simple chat to agentic AI (agents spawning subagents to execute complex tasks) has exploded energy use in ways the industry hasn't disclosed
- His personal heavy usage still only matched running an electric dryer annually, but multiply that across millions of users and you see the grid problem forming

### The Signal

Google said a [Gemini](https://wire.fourthweb.ai/tag/google-ai/) text prompt uses 0.24 watt-hours. Sam Altman pegged ChatGPT at 0.34 watt-hours. Both figures date to 2024, when most people were still typing questions into a chat box and getting answers back. That world is gone. The dominant pattern now is [agentic AI](https://www.fastcompany.com/91600746/how-much-energy-does-agentic-ai-actually-use-one-scientist-tracked-every-prompt-he-sent?partner=rss&utm%5Fsource=rss&utm%5Fmedium=feed&utm%5Fcampaign=rss+fastcompany&utm%5Fcontent=rss): you give Claude or GPT a complex instruction, it spins up multiple subagents, each one chewing through its own inference cycles, and the whole operation balloons into something 600 times more expensive than the toy examples the industry gave us.

Hausfather's method was straightforward. He exported his Claude Code history, looked at token usage (the chunks of text these models process), and estimated energy draw based on the correlation between tokens and [compute](https://wire.fourthweb.ai/tag/ai-infrastructure/) cost. Energy is a major part of datacenter operating expenses, so token pricing roughly maps to kilowatt-hours. It's not perfect, but it's more honest than anything [Anthropic](https://wire.fourthweb.ai/tag/anthropic/) or OpenAI have published lately.

> "In the AI world, a year is an eternity. Now you give instructions to an agent to go spin up 20 different subagents doing some massive process that takes orders of magnitude more energy."

The median prompt in his sample hit 150 watt-hours. That's not catastrophic at the individual level. Hausfather's eight weeks of heavy agentic AI use added up to about what an electric dryer burns in a year. But he's one person. Scale that to every knowledge worker, every developer, every analyst who's replaced repetitive work with agent loops, and you start to see why [Microsoft](https://wire.fourthweb.ai/tag/microsoft/) and Google are signing 20-year nuclear deals and why Sam Altman is pitching a $7 trillion chip fabrication plan. The infrastructure gap isn't theoretical. It's here.

**Key differences between then and now:**

- 2024: Simple Q&A chatbot queries, \~0.3 watt-hours per prompt
- 2025: Agentic workflows spawning subprocesses, \~150 watt-hours per median prompt
- That's a 500-600x jump in a single year, with zero public disclosure from the companies driving the shift

The silence from AI labs is the real tell. We have one-year-old estimates from Google and OpenAI about chatbot queries, but nothing about agents. No updated figures. No transparency about what happens when Claude Sonnet 3.5 spins up a coding task that takes 40 minutes and burns through hundreds of thousands of tokens. The industry moved the product, but didn't move the disclosure.

This matters because we're building energy infrastructure based on outdated math. Every grid expansion plan, every nuclear restart, every natural gas peaker plant contract is pricing in the old model. If agentic AI is now the default and it's 600x more expensive, then every forecast about AI's grid impact is off by two orders of magnitude. That's not a rounding error. That's a different future.

### The Implication

If you're building in the agent space, your users' energy costs are about to become your pricing problem. The gap between what people think AI costs and what it actually costs is about to close, and it won't close gently. Expect pressure on inference pricing, expect users to start caring about efficiency the way they care about latency, and expect the first wave of "low-energy agent" startups to launch by end of year.

For everyone else, watch the datacenter deals. When you see a hyperscaler sign a 10-year power contract, check the megawatts. Those numbers will tell you more about where AI is actually going than any product demo. The energy tells the truth.

### Sources

[Fast Company Tech](https://www.fastcompany.com/91600746/how-much-energy-does-agentic-ai-actually-use-one-scientist-tracked-every-prompt-he-sent?partner=rss&utm%5Fsource=rss&utm%5Fmedium=feed&utm%5Fcampaign=rss+fastcompany&utm%5Fcontent=rss)