The bill that made Sarthak Dhawan’s startup briefly famous wasn’t the result of reckless spending. It came from a single setting nobody remembered to switch off.
Dhawan, 21, co-founded Turbo AI, formerly Turbolearn, an app that turns lecture recordings and documents into notes, flashcards, and quizzes, with childhood friend Rudy Arora after both left college, Dhawan from Northwestern, Arora from Duke, at the end of their sophomore year.
This year, the company’s Claude Code bill hit roughly $30,000 in a single month, well above its usual $20,000 baseline, and Dhawan has been telling the story publicly not as a cautionary tale, but as one he’d repeat.
What Actually Caused the Bill, and What Changed After
The overspend traced back to Claude Code’s “fast mode,” a setting that trades higher token consumption for lower response latency, which stayed switched on for the team without anyone flagging the cost tradeoff.
Once identified, the fix was straightforward rather than drastic: Turbo AI now defaults to standard mode and reserves fast mode for situations where the latency actually matters, routes routine tasks to smaller, cheaper models, and avoids dumping entire codebases into a model’s context window when a narrower slice would do.
None of those changes required cutting back on how much the team builds, only being more deliberate about which setting a given task actually needs.
Why He Still Calls It Worth It, and Where He’s Not Alone
Dhawan has framed the tradeoff plainly: slowing shipping speed to manage a few thousand extra dollars in token costs would have cost the company more in lost momentum than the bill itself.
He’s been candid about the downside too, admitting his own coding skills are atrophying as more of his day shifts from writing code to planning systems and reviewing what Claude generates, a shift he says every engineer leaning this heavily on AI coding tools is quietly experiencing.
The results so far back the bet: Turbo AI has crossed $13 million in lifetime revenue and roughly 5 million users on a 10-person team, having raised just $750,000 in outside funding and stayed cash-flow positive throughout, with a user base that’s broadened well past students to include professionals at firms like Goldman Sachs and McKinsey.
Not every founder is reaching the same conclusion, though. Pylon CEO Marty Kausas introduced token spending controls after his own company’s Claude usage climbed sharply, and larger organizations including Coinbase and Deloitte have reportedly rolled out AI usage limits of their own.
The industry-wide practice some have started calling “token-maxxing,” using the most capable models as often as possible regardless of cost, is getting a harder look across the board even as founders like Dhawan argue the real risk isn’t spending too much on AI, but spending on the wrong setting for the task.
Source: The Times of India, "A 21-Year-Old Founder Accidentally Spent $30,000 on AI Tokens in a Month. Here's Why He Says It Was Worth Every Dollar"
