The startup accelerator chief calls his heavy Claude Code habit an investment and has open-sourced the toolkit behind it
Garry Tan, the president and CEO of Y Combinator, has put a price tag on his relationship with AI agents. He estimates it at roughly the cost of a senior engineer’s salary.
In a conversation with Isaacson, Tan said he spends “about maybe $100[,000]-$200,000 a year on tokens.” For most people, the monthly AI bill is a chatbot subscription. Tan is playing a different sport.
What Tan is actually paying for
The tokens in question are usage credits for AI models. They are the units developers are billed in when they run large language models through tools like Anthropic’s Claude Code.
Put simply, every prompt, every line of generated code, and every agent task consumes tokens. Run enough agents across enough projects and the meter spins fast.
Tan leans on agents for a wide range of work projects and coding tasks. He has gone further than most executives in describing what that looks like day to day.
He has, by his own joking admission, suffered from “cyber psychosis.” He has also said he sometimes barely sleeps because working with his agents excites him that much.
The figure has shifted somewhat depending on the interview. Across conversations in 2026, Tan has put his annual spending on AI workflows somewhere between $50,000 and $200,000. The most recent estimate sits at the top of that range.
He has a name for the approach: “tokenmaxxing.” The idea is that pouring resources into AI usage lets a small team, or even a single person, produce output that would normally require a much larger organization. He reportedly made it the focus of an August Startup School keynote.
Gstack and the five-day blog
Tan has not kept his methods private. He shared a set of 23 Claude Code skills, bundled under the name gstack, on GitHub.
He has described the release as an effort to help teach others what agents are capable of. The project offers role-based skills meant to simulate an engineering team running inside Claude Code.
Developers appear to have noticed. The gstack repository has gathered between 100,000 and 136,000 stars on GitHub, a sign of broad interest among people who build software for a living.
His most concrete example of the payoff is Garry’s List, a refreshed blogging platform. Tan used Claude Code agents to build it in about five days, at a cost of approximately $200 in tokens.
That number is useful context for the bigger figure. A $200 project is a weekend experiment. A six-figure annual budget implies hundreds of those experiments, or a handful of much larger ones, running more or less constantly.
Tan has also taken the message on the podcast circuit in 2026, including appearances on a16z and YC shows. In those conversations, he has argued that AI-driven workflows can produce hundreds of thousands of lines of code.
Why the head of YC spending like this matters
On its own, one executive’s software bill is a curiosity. Coming from the head of Y Combinator, it reads more like a signal.
YC sits at the front door of the startup economy. Its leader openly treating AI spending as a core operating cost sends a message to the founders who pass through its program and the investors who fund them afterward.
The implicit pitch is a reshuffled cost structure. If one person with a well-configured set of agents can do the work of a small engineering team, the math of early-stage company building changes. A founder might spend on tokens what they once spent on early hires.
There is also a dependency question. A company whose productivity rests on heavy AI usage is exposed to the pricing, availability, and policies of whichever model provider it relies on. In Tan’s case, his public toolkit is built around Anthropic’s Claude Code, so a lot of eggs sit in one basket.
