Garry Tan, the CEO of Y Combinator, spends somewhere between 100,000 and 200,000 dollars a year on tokens for AI agents — not on software licences, but on the literal model tokens that power the agents he works with around the clock. The figure dropped on Walter Isaacson’s podcast The Builders, and it lands at exactly the moment the industry keeps circling the same two facts: agentic AI already works astonishingly well, and it is quietly turning into a luxury good.
This piece unpacks what that spend actually buys, why Garry Tan refuses to treat AI agents like traditional software, what Y Combinator’s internal agent Stray Light says about company-scale context, and what the price tag means for everyone budgeting agentic work on a much smaller salary.
- Garry Tan Puts A Price Tag On His AI Token Habit
- Inside Garry Tan’s gstack: 23 Claude Code Skills
- The Jiminy Cricket Agent: AI That Knows Everything About You
- Stray Light: Y Combinator’s Company Brain
- Why Treating AI Agents Like Traditional Software Fails
- Garry Tan’s Brutal Truth: A Crazy Privilege
- Free Agents From Big Tech Are Not The Answer
- What Garry Tan’s Token Bill Means For Your AI Budget
- References
Garry Tan Puts A Price Tag On His AI Token Habit
The Builders Podcast Confession
Tan has been an evangelist for his agent stack for years, but this time he attached a number to it. “I spend about maybe 100 [thousand] to 200,000 dollars a year on tokens,” Garry Tan told Walter Isaacson on The Builders, the podcast Isaacson records out of Kaleidoscope in New York. The episode, which publishes Tuesday, frames the Y Combinator boss as the archetypal power user: someone for whom agents are no longer an experiment but an operating system for daily work.
None of this should surprise anyone who followed the earlier coverage of his Claude Code setup. Back in March, TechCrunch detailed the love-and-hate around that workflow, including Tan’s own joke that he had caught a case of “cyber psychosis” — staying up late because he was too excited to see what his agents would do next.
What The Bill Looks Like At Full Tilt
Annualised figures hide the texture, so convert them. At the top of his range, Garry Tan’s token habit works out to roughly 16,667 dollars a month, a little under 3,850 dollars a week, or about 548 dollars every single day — before anyone on his team spends a cent. Seen that way, his habit resembles a small SaaS contract, signed by one person, renewing daily.
For perspective, a mid-level engineer at a major tech company costs far more than that, so the argument that agents replace expensive humans starts to look less absurd when one executive gets this much output from text prompts and tool calls. What does look extreme — and Tan says so himself — is that this level of access is simply unaffordable at today’s prices.
Inside Garry Tan's gstack: 23 Claude Code Skills
What A Claude Code Skill Is
The stack Garry Tan leans on is built around Claude Code, Anthropic’s terminal-based coding agent, extended with what he calls skills: reusable folders of instructions, scripts and context that teach the agent how to perform a specific job — review a pull request the way the team likes it, run a launch checklist, or prep a meeting. His public collection, hosted on GitHub under the name gstack, contains 23 of these skills and is free for anyone to copy.
Skills matter because raw agents are impressively capable and completely inconsistent. A skill is how practitioners bolt repeatable behaviour onto something that is allergic to repetition: the agent still improvises, but inside rails a human designed. That distinction — process as code, but executed by a human-like model — is the real unit of work in agentic engineering today.
Why Open-Sourcing The Stack Matters
Tan says he shares gstack to “teach others what agents can do,” which is a strategic move as much as a generous one. He has been loud about wanting American labs to compete on open weights too — this same site covered his call for US labs to distill frontier models — and an openly documented workflow raises the floor for every developer evaluating whether agentic coding is worth the token bill.
It also reveals how a power user thinks about moats. If your advantage comes from hoarding prompts, it evaporates the moment a vendor ships better defaults. If it comes from curating skills, context and judgment about where to point an agent, it compounds. The gstack repo is one engineer’s working assumption about which of those is durable.
The Jiminy Cricket Agent: AI That Knows Everything About You
Twenty Years Of Therapy Notes Included
The most striking number in the interview is not the money. It is the scope of access. “It’s read all of my email and all of my Dropbox. It’s read, like, all of my Apple Notes, even my therapy notes going back 20 years,” Garry Tan said, describing the context his personal agent carries. That is not a chatbot with a text box; that is a model with a life’s worth of memory wired into it.
Most organisations and consumers have not caught up to what that implies. Just last week, Bank of America told customers not to share passwords with AI assistants, a small warning that hints at the compliance storm brewing once agents start routinely holding inbox-, drive- and health-level context. The privacy question for 2027 is not whether agents will read your files — some owners already let them.
Seeing Around Corners
Tan’s own metaphor for the arrangement is Disney’s conscience-on-a-shoulder: “Think of it as a Jiminy Cricket on your shoulder at all times… it knows you, and it’s there to help.” Crucially, he emphasises proactivity: “it’ll actually sort of run ahead and say, ‘What about this?’ Or it’ll help you see around corners.” The value is not answering questions faster; it is an entity that has read everything you have ever written and starts volunteering thoughts before you ask.
That framing also explains the spending. An agent that watches your inbox, drafts replies, preps meetings and nags about the things you forgot is calling the model hundreds of times a day, on long contexts every time. The bill is the price of ambient intelligence, not of occasional answers.
Stray Light: Y Combinator's Company Brain
Total Context As A Moat
Inside Y Combinator, the same philosophy runs at organisational scale. “We now have an internal AI agent, for instance, named Stray Light that kind of has a personality, and then it has a company brain,” Garry Tan explained. “It actually has the total context of every founder we’ve ever backed, anyone who’s ever emailed us, like all the applications.” An accelerator’s entire institutional memory, addressable in natural language, is a concrete capability no competitor can duplicate by scraping the web.
Tan says the agent is “rebuilding the way we operate” at YC — triaging the firehose of startup applications, surfacing relevant alumni context, and doing the connective-tissue work that humans used to do from memory. When batch application season hits, the firm’s bottleneck stops being how much its staff can remember and becomes how well they can prompt.
Personality On Purpose
Note the deliberate oddity: Stray Light has a personality. That choice bothers some enterprise buyers, and it is exactly the point. Tan believes personality — moods, preferences, a consistent character — is a feature that makes people trust, remember and work with agents better. A company brain you never want to talk to is a company brain nobody uses.
It is also a preview of how the technology gets bought. The pitch is no longer “one chat window, many users”; it is “one agent, your whole history.” Once an organisation’s brain lives inside an agent, switching costs look less like a migration project and more like moving house — every drawer in every room has to be unpacked and remembered.
Why Treating AI Agents Like Traditional Software Fails
The Repeatability Trade-Off
As an early adopter of the open-source personal-agent framework OpenClaw and a heavy Claude Code user, Garry Tan has landed on a position that will sound heretical to procurement teams: do not manage agents like software. Some people are so worried about “repeatability and safety that they actually trade off all of the beautiful parts of working with an AI agent, which basically can be so similar to a human,” he argued on the podcast.
| Dimension | Traditional software | AI agents (Tan’s view) |
|---|---|---|
| Behaviour | Deterministic, same input same output | Human-like variance, improvises |
| Failures | Bugs, reproducible stack traces | Misunderstandings, corrected in conversation |
| Improvement | Ship a new version | Coach it, add context, add skills |
| Interface | Menus, forms, fixed features | Goals, files, tools, judgment |
| Billing | Per-seat licence, fixed | Per-token consumption, elastic |
| Governance | Change control, QA gates | Supervision, caps, review of outputs |
The Power-User Mindset
The practical reading: give agents reach, watch them closely, and accept variance the way you accept it from a talented colleague. Tan treats his token budget as R&D, not IT spend; the point of letting the model improvise is precisely that it sometimes finds the move nobody scripted. Traditional governance — approvals for every action, deterministic outputs, per-feature contracts — buys certainty at the cost of the very capability you paid for.
Skeptics will hear “move fast and trust the model” and think of every sloppy AI rollout they have seen. Tan’s counter would be that those failures came from the opposite error: forcing a human-like collaborator into software-shaped boxes, then blaming the technology when it sulked inside them.
What Garry Tan’s Software Rule Means For Product Roadmaps
Garry Tan’s sharpest sentence is also his most practical: agents are not software, so teams should not manage them like software. Traditional roadmaps assume deterministic builds — requirements go in, features come out, milestones close. Garry Tan argues agents arrive as a raw capability you coax, constrain, and re-prompt until the behavior lands inside your tolerance, which makes a Gantt chart feel like theater.
Teams that internalize this stop asking whether an agent feature is “done” and start asking whether it is reliable enough at its current price. Garry Tan frames that trade-off as repeatability versus the parts of the work worth automating at all. For most product organizations the vocabulary shift alone is useful: fewer launches declared, more evaluations run, and an explicit accuracy bar written down before anyone ships.
It also explains why Garry Tan keeps every agent close and names them. If quality is emergent rather than compiled, supervision is the job, not a side quest. His answer to people who dread the babysitting is characteristically blunt: the babysitting is the product.
Garry Tan's Brutal Truth: A Crazy Privilege
Too Expensive To Go Mainstream
For all the enthusiasm, Garry Tan is blunt about the barrier: personal agents will only become mainstream when the price collapses. “It’s too expensive right now,” he said, adding that he knows the amount he spends on AI agents is “a crazy privilege that not everyone can do.” That is one of the clearest statements yet from a tier-one figure that current token economics do not scale to consumers.
And the macro picture gives no comfort — this month’s reporting on cracks in the AI spending boom shows hyperscalers rethinking capital freedom even as inference demand climbs. Vendors are meanwhile shipping better visibility tools rather than cheaper bills: Anthropic just launched spending reports and pooled budgets so companies can see what their AI money buys. Agent suites with per-seat credit markups, like StellarFirm’s much-criticised model, sit on top of an already expensive base.
| User profile | Typical setup | Approx. monthly outlay |
|---|---|---|
| Casual chat user | Mainstream chat subscription | ~20 dollars |
| Committed subscriber | Top consumer tier (e.g. Claude Max style) | ~200 dollars |
| Power developer | Coding agents plus API usage | ~500 dollars |
| Garry Tan | Ambient agents, full life context, OpenClaw plus skills | 8,300–16,667 dollars |
The ladder below shows why his number shocks: even at the low end of his range, Tan’s top-end spend is roughly eighty times the most expensive mainstream subscription.
Every serious prediction about AI becoming boring assumes this ladder collapses — subscriptions, models and context all falling in price at once. Tan’s confession is useful precisely because it puts today’s rung spacing on the record.
The Garry Tan Privilege Debate In Miniature
Comment threads love to read the number as flexing, but Garry Tan’s own framing is closer to an apology: he calls the spend a crazy privilege and says openly that personal agents stay niche until the unit economics collapse. The candor matters because Garry Tan is describing a personal-AI winter that price, not capability, is currently causing.
Skeptics answer that a Y Combinator CEO optimizing his own calendar with agents is not evidence about the other eight billion people. Garry Tan would largely agree; his claim is not equality but direction of travel. He is stress-testing, at his own expense, tools the rest of the market will buy at list price once inference keeps getting cheaper.
There is also a quieter signal inside the admission. When Garry Tan volunteers that the privilege is crazy, he is effectively telling enterprise buyers that today’s agent pricing is still early-adopter pricing. Budget owners hear that clearly: the pilot tax is real, and the discount cycle has not started yet.
Free Agents From Big Tech Are Not The Answer
The Algorithm Problem
If agents are too expensive, why not let a platform give them away? Tan pushes back directly, using Meta as the cautionary tale. Big platforms “control the algorithm, and we all wish that they didn’t at some level,” Garry Tan observed — a free agent subsidised by a social giant inherits the incentives of its sponsor, including everything the engagement economy taught us to distrust.
It is the same tension playing out in models, where Meta’s latest “open” release sparks the question of what open actually means once your assistant, memory and actions all live inside someone else’s stack. A free personal AI is only free if you own the context, the weights and the kill switch.
A Personal AI You Program Yourself
The alternative Tan sketches is disarmingly concrete: “you could imagine a personal AI that you control, that you program.” That sentence describes his own reality today — self-hosted skills, files on his own drives, a GitHub repo anyone can fork. What he calls crazy privilege is also a blueprint: the shape of personal AI is already visible; it just costs too much to copy.
Notice who benefits if prices fall and ownership stays with users: independent labs, local-run models, repairable setups, and developers selling skills instead of renting seats. Tan’s interview is effectively a manifesto for that future — he has simply arrived there first, at a price he openly calls absurd.
What Garry Tan's Token Bill Means For Your AI Budget
Rationing Tokens Without Losing The Magic
You do not need 200,000 dollars to learn from this. The lesson is architectural: agents pay rent. Context is the biggest recurring cost, so curate what your agent reads before wiring it to everything — exactly as Garry Tan’s multi-decade access shows the risk of not deciding. Pick the billing model that matches how you work, cap it, and review monthly.
| Budget model | How it works | Best for | Watch out |
|---|---|---|---|
| Flat subscription | One monthly fee, rate limits | Predictable personal use | Limits throttled at worst time |
| API consumption | Pay per million tokens | Bursty deep work | One runaway agent, one nasty invoice |
| Hybrid seat-plus-usage | Base fee plus metered extras | Teams with uneven demand | Credit markups hidden in plan text |
| Self-hosted weights | Own hardware or rented GPUs | Privacy and steady heavy use | You maintain everything |
Three Signals To Watch
First, per-token prices keep falling faster than capabilities disappoint; the history of compute is on Tan’s side. Second, context efficiency is the new model race — memory summaries, cached prompts and distilled skills like gstack cut repeat spend dramatically. Third, budget tooling is standardising: pooled wallets, per-agent caps and spend dashboards are landing now, which tells you enterprise demand for rationing is real.
The personal, self-programmable agent is coming to normal households; in Tan’s telling the only question is how fast the bill drops to a normal household’s price. Until then, his 200,000-dollar confession is the most honest benchmark of what agentic AI costs when someone refuses to ration it — and a preview of the budget line your own successor will not understand.
Copying Garry Tan’s Playbook On A Real Budget
You can steal Garry Tan’s structure without his payroll. First, he pays for constant use rather than occasional demos: agents live where the work happens, so tokens buy output instead of novelty. Second, Garry Tan splits spend by job, putting a cheap model on rote tasks and an expensive one only where judgment pays. Third, everything runs inside skills he can read and audit.
The discipline is what makes the bill defensible. A 3,000-dollar month that replaces contractor invoices is a different conversation from a 3,000-dollar month of prompt play. Garry Tan can absorb 200,000 dollars a year; you cannot, but you can copy the ledger habit — track tokens the way SaaS teams track cloud spend, per feature, per team, per outcome. The real Garry Tan lesson is accounting, not extravagance.
Start where Garry Tan started: one agent, one job, one measurable week. If the tokens produce something you would otherwise pay a human to do, widen the budget. If they do not, the future he predicts — where this level of access is normal — has not arrived for that particular task yet.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
