Garry Tan’s AI token bill shows how expensive agent-heavy startups are becoming

Y Combinator chief executive Garry Tan says he is spending as much as $200,000 a year on AI tokens and is urging founders to embrace heavy agent use as a serious operating practice, not a novelty. His comments, reported across TechCrunch, Crypto Briefing and AOL.com in early November 2026, amount to a public argument that AI work can justify real budget line items — even if the approach is still far too costly for most people.

What Tan is saying about AI spending

Tan’s central message is simple: if founders want to move faster with AI agents, they should stop treating token usage like an expense to minimize at all costs. According to TechCrunch, he said he spends about $100,000 to $200,000 a year on tokens, describing the outlay as covering work projects and coding tasks. Crypto Briefing reported a broader range across 2026 interviews, putting his annual AI spend between $50,000 and $200,000.

That gap matters less as a contradiction than as a sign of how Tan talks about the subject in different settings. The number is large either way, and it is meant to be. In his view, heavy use of model calls, tools and agents is not waste; it is a way to buy speed, leverage and output.

“I spend about maybe $100[,000] to $200,000 a year on tokens.”

That quote, first reported by TechCrunch, captures the scale of the commitment. Tan framed the spending as something founders should think about more like payroll or infrastructure than like a discretionary software subscription.

Why does Tan think founders should use more tokens?

Because, he argues, the payoff is in workflow compression. On an A16z podcast released on November 2, according to AOL.com, Tan said founders should “tune it all the way up” and load in very large token counts. AOL.com also reported that he told listeners heavy agent use could let founders “live in 2028,” his shorthand for being effectively ahead of the current software curve.

In practical terms, Tan’s pitch is that founders should use agents aggressively for coding, project support and task execution instead of limiting them to occasional prompts. AOL.com reported that he estimated full-strength agent use could cost $50,000 to $100,000 a year and suggested a markdown file can serve as an employee-like instruction set for an AI agent.

That philosophy is a direct challenge to the instinct many teams still have: keep the bill low and the usage contained. Tan’s view flips that logic. If the model can do more work, then using more of it may be the rational move.

How Tan uses AI in his own work

Tan has described a personal workflow that goes far beyond casual chatbot use. According to TechCrunch, he said his agents have read his email, Dropbox, Apple Notes and even therapy notes spanning 20 years. He also likened a personal AI agent to a constant helper that can anticipate needs and act ahead.

That is a striking description because it positions the agent less like a search box and more like a delegated assistant with memory, context and initiative. The implication is that the value of the system comes from access to broad personal information and the ability to act across it, not just from generating text on demand.

Tan has also pointed to more concrete tooling. TechCrunch reported that he uses Claude Code heavily for agents and shared 23 skills on GitHub as “gstack.” Crypto Briefing said gstack has drawn roughly 100,000 to 136,000 GitHub stars, and that Tan built Garry’s List with Claude Code agents in about five days for about $200 in tokens.

If that claim holds up, it is the kind of example founders notice. A fast build at a relatively small direct token cost is a strong demonstration of the upside Tan is selling: spend more on the machine, and ship faster with fewer human hours.

What is YC doing with agents internally?

Y Combinator is not just preaching the agent message from the outside. TechCrunch reported that YC is using an internal agent called Stray Light to help rebuild company operations. According to that report, Stray Light has a personality and access to founder and applicant context at YC, suggesting it is being used as more than a narrow automation tool.

That detail matters because YC is one of the most influential startup organizations in Silicon Valley. When its chief executive describes internal agent use as part of operational rebuilding, it signals to founders that this is not a fringe experiment. It is something the accelerator itself is testing in its own house.

Still, the precise scope is unclear. The fact sheet does not say how broadly Stray Light is being used, whether it is in full deployment or early testing, or how much of YC’s workflow has been shifted onto the system. Those are the questions that will determine whether this becomes a model for other startup organizations or remains a high-profile pilot.

Why it matters

Tan’s comments matter because they turn AI usage into a budget and strategy question, not just a product question. If a top YC executive is willing to say openly that he spends six figures a year on tokens, that changes the tone around what serious AI use looks like for startups.

For founders, the message cuts both ways. On one hand, Tan is making the case that agents can materially accelerate coding, operations and research. On the other hand, he is also signaling that the current version of this workflow is expensive enough to favor well-funded teams, which could widen the gap between startups that can afford heavy AI usage and those that cannot.

For model providers, the remarks reinforce an emerging business reality: the people building the most ambitious AI workflows may become major customers. That is good for revenue, but it also creates dependence. If token costs remain high, founders may be locked into whichever provider or stack they can afford to use at scale.

For readers watching the market, the broader takeaway is that AI agents are moving from demo culture into line-item culture. The conversation is shifting from “Can the model do this?” to “What does it cost to have the model do this all day?”

How the numbers compare

Different reports in early November 2026 put Tan’s AI spending in a slightly different frame, but the broad picture is the same: he spends a lot, and he wants others to think bigger about token use.

Background: how tokenmaxxing became a talking point

The backdrop here is the rise of AI agents, software systems that can do more than answer a prompt. They can take actions, keep context and chain together tasks. That means token usage can climb quickly, because an agent can consume large amounts of model capacity while it reasons through a problem, accesses files or iterates on code.

Tokenmaxxing is the term that has emerged for deliberately using large numbers of model tokens to maximize output rather than minimize cost. According to the fact sheet, Tan turned tokenmaxxing into a Startup School keynote theme in August 2026. Crypto Briefing said that by then the idea was already part of his public platform, and the later podcast appearances made the message even louder.

That timing matters because the debate is shifting. On one side are builders like Tan, who see larger token budgets as a productivity advantage. On the other are executives who argue the behavior is fading. AOL.com reported that Uber tech chief Praveen Neppalli Naga recently argued tokenmaxxing is ending, and that Cognition CEO Scott Wu had previously criticized it as overdone.

So the argument is not just about how much to spend. It is about whether the entire “spend more to move faster” mindset is a temporary phase or the foundation of a new operational norm.

What Tan is pushing back against

Tan is pushing back against the idea that efficient AI use means always keeping usage low. He wants user-controlled personal AI, not free agents run by Big Tech, according to TechCrunch. That distinction suggests a broader concern about who owns the assistant, who sets its boundaries and whose interests it serves.

He has also warned that current agent costs are too high for most people. That warning makes his own behavior look less like a prescription for everybody and more like a preview of what frontier usage looks like before prices fall.

What differs across the reporting

The major outlets are aligned on the core story, but they emphasize slightly different angles. TechCrunch focused on Tan’s annual token spend, his personal agent setup and YC’s internal tool, Stray Light. Crypto Briefing highlighted his spending range across 2026 interviews, the gstack GitHub project and the Garry’s List example. AOL.com emphasized the podcast framing, the “live in 2028” language and the token-volume advice.

Those differences do not change the substance. They do, however, show how Tan’s message is traveling through different parts of the tech conversation: startup operations, developer tooling and the culture war around how aggressively to use AI.

What happens next

Several questions remain open. The fact sheet does not say how much of Tan’s token use is personal versus YC-related, how long YC plans to keep Stray Light in use, or whether the internal agent will expand beyond early testing. It also does not specify which AI model providers Tan and YC rely on most over time.

The biggest near-term question is whether falling prices will make Tan’s style of agent use more common. If token costs decline materially, the strategy could spread beyond elite startups. If not, heavy agent use may remain concentrated among teams with substantial resources and a willingness to spend aggressively on speed.

Another point to watch is whether other startup leaders follow Tan’s lead or keep pushing in the opposite direction. If more executives begin treating large token budgets as routine operating expense, the market may move toward agent-heavy workflows faster than many people expect. If not, Tan’s stance may remain a marker of a particularly high-spend corner of startup culture.

Timeline

Here is the sequence of events and reports that brought Tan’s remarks into focus:

• August 2026: Crypto Briefing says Tan made tokenmaxxing a Startup School keynote theme.
• November 2, 2026: AOL.com says an A16z podcast episode featuring Tan was released.
• November 3, 2026: TechCrunch published its report on Tan’s $100,000 to $200,000 token spend.
• November 3, 2026: Crypto Briefing published its own report on Tan’s AI spending and related projects.
• Tuesday: TechCrunch said a Walter Isaacson podcast episode featuring Tan was set to publish.

The bigger signal for startups

Tan’s public stance is ultimately a signal about how a segment of Silicon Valley now thinks about AI labor. The most ambitious users are no longer just asking whether an agent can answer a question. They are asking whether it can sit inside the workflow, read the history, anticipate the next step and keep going long enough to make human time stretch further.

That is why the token bill is the story. It shows the operational cost of treating AI as a persistent worker rather than a sporadic tool. And it suggests that for some founders, the new measure of ambition may not be how little they spend on AI — but how much they are willing to spend to make it feel like a real part of the company.