Garry Tan, head of startup accelerator Y Combinator, estimates that he spends $100,000 to $200,000 a year on tokens for AI agents. He uses them for coding, interview preparation and everyday tasks. Tan acknowledges that a budget like this is a privilege that not everyone can afford.
The entrepreneur discussed his setup with Walter Isaacson on The Builders, as reported by TechCrunch. The amount is Tan's rough estimate, not a published invoice or the typical cost of an agent. The episode description says the interview was recorded on June 19, 2026.
Why agents use so many tokens
Tokens are small pieces of text that a model processes as input or creates in an answer. An agent may call a model several times to solve one task: read documents, make a plan, inspect a tool's output and continue working. Each step adds more data to process and can increase the bill when a service charges by token.
The cost of a particular task depends on the model, the amount of context and the number of calls. Tan's estimate doesn't tell us the price of a single request or the budget someone else needs. It describes unusually intensive personal use, not a minimum required to get started.
According to Tan, his agents help with coding and other projects. One studied Isaacson's podcast before the recording to help him prepare for potential questions. That's a familiar research workflow: a person sets the goal and then checks whether the material is useful.
An assistant with personal context
The setup goes beyond interview preparation. TechCrunch reports that Tan said his agent had read his email, Dropbox files and Apple Notes, including 20 years of therapy notes. His idea is that an assistant with this context can suggest ideas and notice things a user hasn't thought to ask about yet.
Broad access may make suggestions more useful, but it also raises questions about trust and privacy. Personal notes, correspondence and work documents contain information about other people too. Before connecting an agent, check where data goes, who can access it and how to disconnect the integration. Tan's account describes his choice, not a recommendation to open your whole digital life to an agent.
An agent inside Y Combinator
Y Combinator uses AI assistance too. Tan says the accelerator has an internal agent called Stray Light with context about founders it has backed, applications and correspondence. He says the system is helping change how the organization works. The TechCrunch report does not establish how well it performs specific tasks or how access is controlled.
Tan has also shared gstack, an open-source collection of tools and instructions for AI-assisted development. It shows one way to organize agent tasks, but it doesn't explain his annual budget by itself: that depends on the models and workload involved.
Who else uses large amounts of AI?
Tan is not alone in discussing intensive model use. These figures aren't directly comparable: some describe personal spending, others a team's workload, and some are valued at public API rates rather than an actual bill.
- Peter Steinberger and the OpenClaw team. In May 2026, OpenClaw's creator shared usage statistics showing about $1.3 million over 30 days, covering roughly 603 billion tokens and 7.6 million requests. Tom's Hardware reported that a three-person team used around 100 Codex instances and OpenAI, Steinberger's employer, covered the cost. The figure reflected fast-mode pricing and was not his personal bill.
- OpenAI researchers. OpenAI reported that by mid-August 2026, the median researcher used over $600 of coding-agent inference per day, while the 90th-percentile figure exceeded $7,000. These are estimates at API prices, not disclosed internal compute costs. The percentiles describe a distribution, not fixed groups of people.
- GenBrain AI. In its own breakdown, the company said seven agents cost about $1,150 a month including infrastructure, with around $710 attributed to Claude tokens. Those are the company's reported costs for the whole system, not a single agent.
The range is wide. Agent count, workload and the way costs are calculated all matter, so comparing totals without those details can be misleading.
Where to start and how to control costs
Tan believes personal agents need to become cheaper before they can reach a wider audience. He calls his own spending a privilege. A large number alone doesn't prove that every token paid for itself: the account includes no independent measurement of time saved or results.
If you want to try AI, start with one clear task, such as reviewing a document or getting help with a piece of code. Set a budget, check the output and then decide whether to expand the workflow. GPTunneL lets you choose models for different tasks and pay from one balance.