Garry Tan says he spends up to $200,000 a year on tokens

Six figures on a software subscription used to buy a company a seat on a sales platform. Garry Tan says it is now what he spends on AI tokens for himself. Tan is the president and CEO of Y Combinator. TechCrunch's Julie Bort reported the number on 2 November 2026. The source was an interview on Walter Isaacson's new podcast called The Builders.

Tan's quoted figure was about $100,000 to $200,000 a year on tokens. TechCrunch wrote the number in that range. Tan called the spend a crazy privilege. He said price is the barrier for everyone else.

What his agent can see

Tan's personal AI agent reads his email. It reads his Dropbox files. It reads his Apple Notes. It also reads his therapy notes. The time span he gave was twenty years. His quote was that the agent has read all of his email and all of his Dropbox. The agent has also read all of his Apple Notes. Those notes include his therapy notes going back twenty years.

He described the shape of the help in one line. He wants the agent to do the kinds of things anyone does in a day. He wants it to do them more autonomously. He called it a Jiminy Cricket on your shoulder at all times.

The YC agent and the public skill pack

Tan also described a Y Combinator in-house agent. The agent is called Stray Light. TechCrunch wrote that the agent has a personality and company context. That context includes the history of founders. It includes their applications and the email across YC-backed companies. The piece does not say how many YC staff use it. It also does not say how often.

Tan has published 23 of his Claude Code skills on GitHub. He calls the pack gstack. TechCrunch frames the release as his way of helping others learn what agents can do.

What this says about the top of the market

The number is one person's spend and nothing more. It is not Y Combinator's budget. It is not a published average. It is not a floor for working with agents. TechCrunch quotes it as Tan's own statement about his own use.

Three things make it worth reading anyway. First, Tan is a buyer who talks to thousands of founders a year. His working setup leaks into every batch of advice. Second, the mechanics are a design choice most teams have not made yet. One agent with access to twenty years of personal data is a precedent other buyers can point at. Third, the price itself is a target. A reader who wants Tan's level of autonomy and spends far less has a gap to measure.

The practical move for a team is less dramatic. Pick one real workflow. Run it against the agent for a month. Watch the token bill. Compare it to what the same hours cost in payroll. If the ratio is bad, cut the workflow or cut the agent. The six-figure headline matters less than the honest number for one job.

Source

Julie Bort, "Y Combinator's Garry Tan uses AI agents so much, he spends up to $200K a year on tokens," TechCrunch. The quoted interview was Walter Isaacson's podcast The Builders.