Y Combinator's Garry Tan Says His AI Agents Cost $100,000 to $200,000 a Year in Tokens

Y Combinator CEO Garry Tan says he spends about $100,000 to $200,000 a year on tokens (the units of text a model reads and writes, billed per unit) to run AI agents, according to TechCrunch, which reported the remarks on 2 November 2026.

Tan made the comment on Walter Isaacson's podcast The Builders, in an episode set to publish on Tuesday, 3 November. He described an agent setup with broad access to his own data. In his account, the agent has read all of his email, his Dropbox, and his Apple Notes, including therapy notes going back 20 years.

Tan also published a set of 23 Claude Code skills, reusable instruction packages that extend Anthropic's coding agent, under the name gstack on GitHub. He said the aim is to show other developers what agents can do in practice. Y Combinator itself runs an internal agent named Stray Light, which Tan said is rebuilding the way YC operates.

On cost, Tan said personal AI agents will become mainstream if prices fall, adding that they are too expensive at present. He described his own level of spending as a “crazy privilege” that not everyone can afford. The remarks echo advice he gave earlier in the year. In August, Business Insider and the Times of India reported that he told founders to give themselves permission to spend heavily on AI tokens despite the cost.

Tan's broader framing has also been public through his Startup School 2026 talk, “Own Your Intelligence,” which addressed what he called the era of personal AGI, in which individuals run agents on their own data.

In my view, the number itself matters less than the arithmetic behind it. A six-figure annual token bill is a plausible line item for a founder running agents continuously across email, documents, and personal notes, with long context windows (the amount of text a model can consider at once) and frequent tool calls. It is not a plausible line item for most knowledge workers. Tan's own framing makes that distinction: the spend is a privilege, and his argument depends on unit prices falling far enough that the same workload becomes ordinary.

The broader context here is a question that every enterprise buyer of inference is already working through. Per-token pricing, caching, model routing, and smaller distilled models all push the cost curve down. Agent architectures push consumption up, because a single task can fan out into many model calls, retrievals, and retries. Whether personal agents become mainstream will depend on which of those forces wins at the consumer price point, not on capability demonstrations alone.

The privacy side of the setup deserves the same attention as the cost. Giving an agent read access to two decades of therapy notes is a design choice with its own threat model. That model covers local versus hosted inference, data retention by model providers, and what happens when credentials or connectors are compromised. Tan's account describes the access. It does not describe the controls around it, so readers evaluating a similar setup will need to answer those questions themselves.

What Tan is describing, at its core, is a shift in where an individual's data and computation sit. An agent that holds a person's full record and acts on it could change how people manage their time, money, and relationships. The price of running one is the constraint he named himself, and it is the one most likely to determine how quickly that change reaches people beyond founders and engineers.