Y Combinator's Garry Tan spends up to $200,000 a year on AI tokens

Y Combinator CEO Garry Tan spends between $100,000 and $200,000 a year on tokens to run his personal AI agents. He said so himself, on a new episode of Walter Isaacson's podcast The Builders, set to publish on Tuesday, and the number puts a hard figure on what deep, daily reliance on AI agents actually costs at the extreme end.

That figure matters beyond one Silicon Valley executive's expense account. Tan runs the startup accelerator that has seeded thousands of companies, and his spending hints at the gap between what AI agents can do today and what ordinary users could ever afford. If a power user burns six figures a year on model access, personal agents remain a luxury product, not a mass-market tool.

Tan hasn't hidden his enthusiasm before. He has jokingly admitted to suffering from "cyber psychosis" and sometimes barely sleeping because he's so excited to be working with his AI agents. On Isaacson's show, that passion finally got a price tag.

"I spend about maybe $100[,000]-$200,000 a year on tokens," Tan told Isaacson.

What does Tan actually use agents for?

The workload splits into two buckets. The first looks familiar: work projects and coding tasks, plus research. He used an agent to prepare for Isaacson's podcast, prepping for the kinds of questions he might be asked. Plenty of professionals now lean on AI that way.

The second bucket goes much further. Tan's agent helps him run his life.

"It's read all of my email and all of my Dropbox. It's read, like, all of my Apple Notes, even my therapy notes going back 20 years," he said.

That level of access is the mechanism behind his framing of agents as companions rather than tools. He describes using AI agents as getting help with "almost anything that you do in your day-to-day life, and for it to do it somewhat more autonomously. Think of it as a Jiminy Cricket on your shoulder at all times. Like, almost anything you want to do, it knows you, and it's there to help. And it will actually sort of run ahead and say, 'What about this?' Or, it'll help you see around corners."

The takeaway: Tan's agents don't just answer questions — they hold the full context of his life and act on it.

Why has he published his agent setup on GitHub?

Tan has freely shared his set of 23 Claude Code skills, called gstack, on GitHub. He frames it as a teaching effort, a way to help others learn what agents can do. For developers watching how serious practitioners configure agent workflows, an open toolkit from YC's CEO is a rare public artifact. It's one thing to hear an executive praise agents; it's another to inspect the 23 skills he actually runs.

The takeaway: the gstack repository turns Tan's advocacy into something others can copy and test.

How does Y Combinator itself use agents?

The usage isn't only personal. Tan said Y Combinator is running an agent that is "rebuilding the way we operate at YC."

"We now have an internal AI agent, for instance, named Stray Light that kind of has a personality, and then it has a company brain. Like it actually has the total context of every founder we've ever backed, anyone who's ever emailed us, like all the applications," he said.

Stray Light's design is worth noting. YC gave the agent a personality and a company brain, then fed it the accelerator's entire institutional memory — every founder backed, every applicant, every email. That's an organizational knowledge base wired directly into an AI system with a degree of autonomy. For a firm whose core asset is pattern recognition across thousands of startups, the stakes of that experiment are obvious.

The takeaway: YC is testing whether an agent holding the firm's total context can reshape how the accelerator operates.

What has Tan learned about how agents should be treated?

As an early OpenClaw adopter who is now a power user, Tan says the biggest lesson is not to treat AI agents like traditional software. He argues that some people are so worried about "repeatability and safety that they actually trade off all of the beautiful parts of working with an AI agent, which basically can be so similar to a human."

That's a contentious position. Traditional software engineering treats predictable, repeatable behavior as a baseline requirement. Tan flips that logic: clamp down too hard on an agent's variability, he suggests, and you lose what makes it useful. It's a live debate in the industry as companies decide how much autonomy to grant agents touching real systems and real data — like, in Tan's case, 20 years of therapy notes.

The takeaway: Tan believes over-engineering for safety strips agents of their human-like value.

Will personal agents go mainstream?

Tan believes personal AI agents will eventually become mainstream, but only if one major problem gets solved: costs have to come down.

"It's too expensive right now," he said. He also acknowledges the privilege embedded in his own usage, saying he knows the amount of money he spends on AI agents is "a crazy privilege that not everyone can do."

This is the tension at the center of the story. Agent capabilities are escalating — Tan's setup researches, codes, manages email, and runs ahead of him with suggestions — while the token bill for that intensity sits at six figures. Mainstream adoption, in Tan's own analysis, waits on cheaper inference, not better marketing.

The takeaway: in Tan's view, price is the single gate between power users and everyone else.

Are free big-tech agents the answer?

Cheaper could mean free, but Tan is wary of that path. When discussing the future he'd like tech to build, and how Meta as a social media company has impacted us, he pointed to the concentration of control.

"This is a huge company that controls the algorithm, and we all wish that they didn't at some level," he said. "So you could imagine a personal AI that you control, that you program."

The implication is a fork in the road for the agent market. One route runs through big-tech platforms offering free agents funded by advertising and algorithmic control — the Meta model applied to a new interface. The other runs through agents users own, program and pay for directly, closer to what Tan does today at a cost most people can't bear. His six-figure token bill is effectively the current price of that independence.

That choice — subsidized agents you don't control versus expensive agents you do — is likely to shape how the personal-agent market develops as costs fall. Tan's bet is on the second path, and his gstack repository is an early blueprint for what building it yourself actually looks like.