My voice note, trimmed and put in order. Facts in the boxes.
The 1-minute version
- The token cost of an AI agent starts before any work: one agent used tokens just to say "OK".
- A small job with four agents cost about tokens. My guess for one assistant doing it alone: about 25,000.
- Anthropic, who make Claude, say multi-agent setups use about the tokens of a normal chat.
- A cheaper team I built on Google's Gemini lost a blind writing test, out of 40.
- My rule now: an agent only for fresh eyes on finished work, or for reading too big for one chat.
Why I wanted an AI team
I wanted an AI team because I'm building a startup on my own. So I thought the fastest way to get a team is to get an AI team to work.
What one agent actually costs
I expected it to cost like normal token usage. I thought it would be about what a session takes when I keep going in it. But just starting a new team member on something costs a lot, 21,000 tokens or so.
20,971tokens
for one agent to reply "OK"
Tokens are how AI tools count your usage, like mobile data. Before it even reads your task, every agent loads its own instructions, its tools and all your saved rules. That's the 20,971. Part of it was my own saved rules, so your number will be different. Measured 8 Oct 2026.
It would take four or five times more to answer it when I'm using the agents, right? So that's why it surprised me a lot.
Tokens used, same day, same plan
I measured the solid bars on 8 Oct 2026. The hatched bar is my guess.
Why I scrapped it
Tokens are limited on the Pro plan, so every token has to count. And I also tried the Gemini thing, but it was not working because the quality was lower. The workflow which I am trying to create is to be the best of its quality with the least token usage.
The Gemini test
- Writer drafts
- Critic and editor read the same file
- Writer merges
- Another AI judges, blind
Gemini team 20/40, one assistant alone 35/40. The judge called the team's version choppy and robotic. One test, so not proof.
When an agent is worth it
My rule now
- Worth it
- Fresh eyes on finished work, the way you'd never proofread your own poster.
- Worth it
- Reading so much that it would clog your main chat.
- Not worth it
- Anything one chat can do.
If you're about to build one
So what would I tell someone about to build the same thing? I would tell them that you should build it. I mean, it's not that there is no solution for it, but it's that I haven't found it.
Before you build an AI team
- Make one agent and ask it to reply "OK". Check how many tokens that used: that's your start-up fee.
- Multiply it by how many agents one job will wake up.
- Run the same job in one normal chat and compare.
- Keep agents only for the two "worth it" cases above.
I am mainly a vibe coder, as people say it nowadays. So I take help of AI as a tool, and I only use my brain to decide what to do. So yeah, you can build the same thing if you can find the solution for it, or wait until the AI agents find one themselves.
For now we just continue normally, like a normal person who uses Claude to build something. The AI team can wait until the numbers work.
Sources
- Anthropic engineering (2025)
- Walden Yan, Cognition (2025)
- Cemri et al., "Why Do Multi-Agent LLM Systems Fail?" (NeurIPS 2025)
- dev.to start-up cost test
How this post was made: the main text is my voice note, trimmed and put in order. Claude wrote the headings and the boxes from the linked sources, Gemini checked the facts, and I read and approved it. Numbers are from my own tests on 8 October 2026. This version of my words was remixed by AI; the numbers and boxes are unchanged.