This told essay is based on a conversation with 21-year-old Sarthak Dhawan. Co-founder of Turbo AIbased in New York. His words have been edited for length and clarity.
There was a month this year where I accidentally spent about $30,000 on Claude Code tokens, and it wasn’t great.
I learned something from that month, but I don’t think it was a complete mistake. Shipments were high this month, and our spending reflected that. A month with a lot of tokens usually means you are innovating or trying something new.
My co-founder Rudy and I launched the AI Learning Tools app in January 2024 while in college, but dropped out last year to pursue the business full-time. In my opinion, our bottleneck is shipping speed, not thousands of dollars of tokens. Had we slowed down to manage token spending, we would have lost even more momentum.
When we started coding the app, the AI code generation model was pretty terrible. I wrote the code myself from scratch. Coding has changed a lot these days, and the actual tasks of engineering feel very different.
AI has changed my job from writing code to reviewing it
I spend much of my day planning and explaining how things work at a high level and reviewing AI-generated code, rather than actually writing the code myself. There are many things to check on the atmosphere as you go along.
I feel like my coding skills are atrophying. I think all engineers who use Claude Code know that it has such an impact on them. The more Claude takes over tasks, the less the engineers actually know about their codebase.
For example, 20 years ago, a group of engineers working on a codebase knew the details of that code deeply. They had a very good understanding of the architectural decisions because they had made all of them themselves.
When the AI writes the code, the code base moves to this entity that the user cannot understand. At the same time, its use cannot be avoided, since productivity will be significantly reduced. The difference in speed is completely incomparable.
AI costs have increased, but no AI spending budget has been set
As AI coding became more useful over time, we did not set a strict budget for spending AI tokens. There is no approval process for token usage, although we loosely monitor it.
We have about 10 people on our team, and costs are increasing as more people use AI for more tasks. As long as you’re driving the output, you’re fine. On average, AI tools for software development cost about $20,000 per month.
As for measuring token cost per developer, we’re just saying use whatever you prefer, and we’ll continue to do so for the time being.
I left one Claude setting on and my bill was almost $30,000, but the fix was easy
In April, the bill for AI tokens jumped to about $30,000. One of the reasons my bill was so high was because I was working in fast mode with Claude without realizing it.
A setting called Claude Code’s fast mode makes it feel faster, but it also costs a lot more per token. When I turned on express mode and left it alone, my bill skyrocketed.
Turn on Fast Mode if latency is currently an issue when pairing. Otherwise, leave it unchecked. There was almost no change in output speed when switching from high speed mode. Even in normal mode, it’s fast enough and the quality is the same, so it was an easy save that didn’t feel too deep.
Continue reading the Tiny Teams series
Don’t think too much about saving tokens
To save tokens, we take the easy way out, like standard mode by default, lightweight models for simple tasks, and not dumping the entire codebase into the context, but we don’t stress every dollar. This year, our lifetime revenue exceeded $13 million.
We still think that if it increases productivity, it’s probably worth it to the business in the long run.
Do you have a story to share about how you are using AI tokens? If so, please contact our reporters. aapplegate@businessinsider.com.