Only if it frees hours you spend on judgment. Garry Tan told Walter Isaacson's podcast (previewed November 2, 2026) he spends $100,000-$200,000 a year on tokens; at $200,000 that must free about four hours a week at a $1,000 hour, while Claude Max at $100 a month needs about one hour a year. Naval Ravikant's leverage rule applies: it multiplies judgment, and busywork too.
A number like that is built to travel. On an episode of Walter Isaacson's podcast The Builders, previewed by TechCrunch on November 2, 2026, Y Combinator CEO Garry Tan said: "I spend about maybe $100,000 to $200,000 a year on tokens." The reactions split between "serious founders spend like this now" and "agents are a rich person's toy." Both skip the useful question: what does the money buy, and does the same purchase make sense at your scale?
Start with what he described. Tan uses agents to prepare for interviews, to write code and to help run his life; he says his agent has "read all of my email and all of my Dropbox," his Apple Notes, and "even my therapy notes going back 20 years." He shares gstack, his set of 23 Claude Code skills, free on GitHub under an MIT license. And he added the caveat most headlines dropped: "It's too expensive right now," and his spend is "a crazy privilege that not everyone can do."
Now put the figure next to ordinary numbers. Anthropic's Claude Max plan costs $100 a month for five times the Pro plan's usage every five hours and every week. The Ultra plan, renamed from Max 20x, costs $200 a month for ten times Pro's weekly usage, with no five-hour limit for verified accounts. On the metered side, Anthropic's own Claude Code cost guide puts the average across enterprise deployments at about $13 per developer per active day, or $150-250 a month, with 90% of users under $30 a day; those are vendor figures, not an audit. Thirteen dollars across 250 working days is about $3,250 a year. Tan's upper figure is roughly 60 times that.
The lens I find most useful is Naval Ravikant's, from The Almanack of Naval Ravikant. Naval calls code permissionless leverage: it works while you sleep and needs nobody's approval. He also argues that in an age of leverage, judgment matters more than effort. Tokens are the purest leverage you can buy, but leverage multiplies whatever you point it at, including busywork. So the honest question is not "how much should I spend?" but "how many hours of my judgment does this free, and do I use them for judgment?"
That gives a break-even test. The table converts each spend level into the hours it must free per year, at two hourly rates. The rates are my assumptions for illustration; swap in your own.
| Annual spend | What it is | Hours it must free at $250/h | Hours it must free at $1,000/h |
|---|---|---|---|
| $1,200 | Claude Max, $100/month | 4.8 h/year | 1.2 h/year |
| $2,400 | Claude Ultra, $200/month | 9.6 h/year | 2.4 h/year |
| $3,250 | Average Claude Code developer (~$13/active day × 250 days) | 13 h/year | 3.25 h/year |
| $100,000 | Tan's lower estimate | 400 h/year (8 h/week) | 100 h/year (2 h/week) |
| $200,000 | Tan's upper estimate | 800 h/year (16 h/week) | 200 h/year (4 h/week) |
The flat plans clear the bar almost by default: an afternoon saved per year pays for them. The bottom rows are different. At $200,000 a year, agents must give back four hours of every working week at a $1,000 hour, or sixteen at a $250 hour. For someone running Y Combinator, that may be a fair trade. For a pre-seed founder it would mean the agents replace most of a full-time employee, which is a claim you should be able to show, not feel.
The second condition is the one people skip: freed hours only count if they go to judgment. ActivTrak's analysis of more than 9,000 employees, reported by Gizmodo on November 3, found AI users' workdays ran 8.4 to 9.1 hours against 7.6 for non-users, and heavy users had shorter uninterrupted stretches than light users. Saved time tends to get refilled with more agent work.
Two more limits. Cost scales with context, because every agent turn re-sends part of what the agent knows; Anthropic's guide says token costs scale with context size, and that a team of agents uses tokens roughly in proportion to its size. Handing an agent twenty years of email and therapy notes is a privacy decision before it is a budget one. And Tan's spend is not evidence that spend causes output; he runs YC with a staff around him.
The rule I take from this week: start on a flat plan, move to metered spend only for a named workflow whose output you can price, and once a month divide your AI bill by the hours it freed. If the result is above your hourly rate, or the freed hours went into more busywork, it is not leverage yet. For the runaway-bill side of the same problem, see how to stop an AI agent from running up a huge bill.
Sources: Y Combinator's Garry Tan uses AI agents so much, he spends up to $200K a year on tokens, TechCrunch, November 2, 2026; garrytan/gstack, GitHub, accessed November 6, 2026; What is the Max plan?, Anthropic Help Center, accessed November 6, 2026; What is the Ultra plan?, Anthropic Help Center, accessed November 6, 2026; Manage costs effectively, Claude Code Docs, Anthropic, accessed November 7, 2026; AI Users Are Working Up to 90 Minutes Longer Every Day, Study Finds, Gizmodo, November 3, 2026; ActivTrak Productivity Lab: AI Users Have Workdays Up to 90 Minutes Longer Than Non-Users, ActivTrak, accessed November 7, 2026.
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · Why Exploration Is Important for Success