A startup accidentally spent $30,000 on AI tokens in a month. Here is what they learned.

A small startup accidentally ran a $30,000 AI token bill in one month and decided it was worth it. The story from Turbo AI co-founder Sarthak Dhawan reveals how easy it is to overspend on AI coding tools, and more importantly, what practical fixes work without killing shipping velocity.

Turbo AI, an AI learning tool with about 10 people, averages around $20,000 per month on AI tooling costs. In April, that bill jumped to roughly $30,000. The cause was not a bigger team or a sudden surge in output. The culprit was Claude Code fast mode, a setting Dhawan left on without realizing how much it cost per token.

What happened

Dhawan, who dropped out of college in 2023 to build Turbo AI with co-founder Rudy, says the team never set a tight budget for AI token spending. There is no formal approval process. Spending is loosely tracked because it is tied to momentum and output. In April, a high-velocity shipping month pushed token consumption higher, and fast mode made each token more expensive.

"When I turned fast mode on and left it on, that's when my bill skyrocketed," Dhawan told Business Insider. The setting made Claude Code feel faster but cost significantly more. The team crossed $13 million in lifetime revenue this year, and Dhawan views the spike as an acceptable cost of fast shipping.

Why AI builders should care

For any team shipping AI products, token costs can creep up fast without a clear signal. The tension is real: faster AI-generated code means more output, but it also creates a less transparent codebase. Dhawan noted that as Claude writes more of the code, engineers lose depth of understanding of the codebase. "The code base shifts into this entity that you don't understand," he said. At the same time, avoiding AI tooling would mean being "way less productive."

This tradeoff matters for founders and developers relying on AI coding tools. The risk is not just cost but also codebase opacity, where architectural decisions become invisible to the team.

Practical implications

Turbo AI found easy wins that reduced token spend without sacrificing output quality. Dhawan now flips fast mode on only when pairing and latency matters. Otherwise, the team keeps it off. "Switching out of fast mode barely made a difference in the speed of output. Normal mode is plenty fast, and the quality's the same," he said.

Other savings include using lighter models for simple tasks and not dumping entire codebases into context. These are straightforward patterns that any AI builder can adopt. The team is not stressing over every dollar, but the cost-aware defaults are worth hundreds or thousands each month.

Caveats

This is a single case study from a very small team (10 people) with a high-revenue product. Token costs and governance practices will vary significantly across organizations and tooling beyond Claude Code frequencies. The evidence comes entirely from one founder's account and may not reflect broader patterns. Teams building on different models or with different shipping cadences should treat these findings as directional, not universal.

FAQs

Why did a startup spend $30,000 on AI tokens in a month?

The spike occurred during a high-velocity shipping month at Turbo AI, when Claude Code fast mode was left on, increasing per-token costs significantly. The team, which normally spends around $20,000 per month, saw the bill jump to roughly $30,000 as a result. The co-founder said fast mode made the tool feel faster but cost more per token.

How can startups balance speed and cost when using AI tooling?

Startups can default to standard mode for routine coding tasks, reserve fast mode only for latency-sensitive pairing, use lighter models for simple tasks, and avoid dumping entire codebases into model context. These are low-effort, high-impact practices. Turbo AI found that switching out of fast mode had no noticeable impact on output speed or quality. The team also keeps a loose eye on spending without a formal approval process.

What is Claude Code fast mode and why does it increase token usage?

Claude Code fast mode is a setting that makes token processing feel faster, but it costs more per token. When left enabled continuously, it can dramatically increase monthly bills. Turbo AI co-founder Sarthak Dhawan discovered that turning it on and leaving it on caused his bill to skyrocket. Now the team uses it only when pairing and latency matters.

What fixes help reduce AI token spending without sacrificing productivity?

Defaulting to standard mode, using lighter models for simple tasks, and avoiding large context dumps are effective, low-friction fixes. Turbo AI implemented these changes and reported no loss in output speed or quality. The team also found that standard mode is plenty fast for most work.