YC's Garry Tan uses AI agents so much, he spends up to $200K a year on tokens
techcrunch.com
techcrunch.com
I wouldn't say this was the norm, but I have burned through £1k of tokens in a couple of days before. And, on that same gig, I realised a test I was considering running would burn £2.5k worth of tokens over the course of a few hours so spent a couple of hours of my own time re-engineering it to get that down to £150.
The point is that I don't necessarily know that spending a shitload of money on tokens really tells us anything useful. If he's getting a return on that investment, fine, whatever.
Obviously most of us couldn't afford to spend that much ourselves, no matter how much value it might deliver, and many of our employers wouldn't be too happy if we spent it on their dime either. Equally, if you can, and it's delivering real value, then I don't see why you wouldn't?
details please, if it's interesting
You can do it if your employer has endless pockets.
Example: $30/1M input tokens (I don’t count output or cache, for simplification), you have several daily agents for everything => 3M tokens daily per agent * 10 agents ~> $900/day => $328k/year (API spending)
And, in fact, I got it from £2500 to £150 by careful model selection (I ran a trial over a small dataset with different models until I found the least expensive/capable model that would reliably do what I needed), using a local model for one step, and paring down the inputs as well so the token count plummeted.
They feel like they're super smart but it's just as smart as hiring a $500k/yr PhD as a PA to schedule your meetings and answer the phone. Sure it does a good job but you're just wasting money and abilities.
They feel super edgy and bleeding edge but all they really do differently is switching the AI chat from auto to the most expensive option all the time. When I see what they're using it for they're only grasping the basics of its abilities.
They're also people sending 2 page AI generated emails that make no sense for literally everything where a "Yeah sure let's do that" would have sufficed.
Unfortunately most of these are business leaders who just want to look edgy and "all in on AI" so we can't really limit their tokens. Yet they're also the ones saying we spend too much and want to limit some of our dev teams actually putting it to good use.
Management succumbed to AI psychosis, and so we were directed to show that we were adopting and using AI, but what we were using it for was not considered important. So we went out of our way to burn as many tokens as possible.
If you tell Claude to be inefficient and burn as many tokens as it can, it’ll happily oblige. It will also offer useful advice as to how to use even more tokens.
Once you factor in cost of electricity and depreciation of hardware, it is more expensive to run locally.
There are many arguments for running local but cost isn't one of them.
Let's do quick back of the napkin math.
- Mac Studio M5 Ultra, 80-core GPU, 256GB memory, 2TB storage -- $11,299.00.
- Power (385w continually 100% of month) is $40/mo (my regional) or $480 annually. Let's say power is 3x more expensive in California, which it probably is, that's still only $1,440 annually in power.
Once you make the jump to agents managing agents the only limit is your budget and your harness's ability to coordinate. We'll all be using far more tokens than this in a year or two.
Or are people paying $1K/mo (subsidized) just because AI is fun? Can you imagine a video game charging $1K/mo?
But you might have hit the nail on the head - yes, it's extremely fun and maybe that's the main reason after all. And FYI many, many people pay a lot more than $1k/m for video games.
In Tan's case, I think the "fun" thing is the right way to look at it. He gets to play at being a big shot developer and show off on Twitter, so clearly he feels he's getting some value from his spending even if no-one else does.
There is a more serious question about to what degree the AI vendors have a duty of care to prevent addicts from getting themselves into financial trouble - perhaps they might in future face the sort of regulation applied to high street bookmakers. I doubt Tan would fail any putative affordability test, though, so it would still be up to him to seek help for his problem if he feels its getting out of control.
It sounds like a lot when I type it like that, but it's not actually that much if you delegate brutally. Like I said, I'm constantly token starved and could do a lot more!
You can get access to DeepSeek 4.1 Flash for almost nothing, on a $1, $10, or $100 subscription. That can complement your 5 Claude accounts, or replace some. DS Flash 4.1 is so fast and cheap it’s ridiculous
This is my daily intrusive thought. "Why the hell am I doing any of this? Why am I not yelling louder at people that have so much and give so little?"
This website is full of people that could save millions of human lives in aggregate, but choose not to and actively argue against the idea because, and I do swear it's a damned curse at this point, "there is no incentive to do it."
I'm not even remotely rich. I have no "savings" or "investments". Yet I have a comfortable life in a house with access to food and care as I need. Sometimes I really hate it.
> “It’s too expensive right now,” he said, adding that he knows the amount of money he spends on AI agents is “a crazy privilege that not everyone can do.”
With AI spend - especially at Anthropic/OpenAI API prices - spending 0.5-1 macbook a month worth of tokens per employee seems normal (from what I've seen, heard, what happens in this forum and thread and what my employer does).
Tokenmaxxxing (followed by commit/PR-maxxxing) is celebrated. Don't think - just launch orchestrator that will create dozen child agent sessions to do useless work to impress your superiors. If buy the end of the day you will ask to increase your token allowance - you're a superhero.
That's hardware.
Sleep is one of the most important things in the world. Why not have the agents work while you sleep?
It was productive, but as you say, it's dangerous to gamble with sleep.
Even better, why not have the agents sleep for you?
EDIT: for those who say this is not realistic, I've had Claude sleep for 10 hours in just 30 seconds. Saved me an entire night.
[1] https://www.amazon.com/Million-Plus-Random-Digits-Symbols-Gu...
People in psychosis don't exactly weigh all of the options with a level head. He may not recognize that he has "cyber psychosis" or that his lack of sleep is a bad thing.
Don’t make me defend Garry Tan…
This is a difficult question to answer ! I would ask a related question which is equally hard. Has AI-use increased your human cognitive capability or decreased it ?
I accept this loss because the productivity is there. & hope to retire in a few years if things don't go south. For people entering the field I worry
what has he produced with that?
I don't think he has produced anything.
Tokenmaxxing was a dumb trend he has not figured out he's out of the loop yet.
Gary spending 250K/year on nothing is not strongly related to pros using $200/plan to do much better.
He's gonna hit the halting problem
Evaluate the original post and response and build a high level learning plan with citations and references that covers each of the topics in the original post. Start the learning plan with an overview of effective prompting techniques with specific examples to research individual topics.
Large language models are usually priced on MTok or “million tokens”. You’ll often see LLMs priced at $X/MTok for input and $Y/MTok for output.
A harness is a program that interfaces with a large language model and provides it tools and context to achieve a goal. Codex and Claude Code are good examples of harnesses.
What exactly all of this accomplishes is not something I can answer for you :)
Tokens are what you get when you have to map a chunk of text to specific numbers, so that the LLM can actually work with it.
https://platform.openai.com/tokenizer
Reducetively speaking:
[Your text] --> [List of numbers (tokens)] --> [List of [List of numbers describing a concept/term/thing/job/item/etc] (embedding)]
That final thing (List of embeddings) is what gets fed into the LLM, so that it can produce (List of embeddings most likely to follow afterwards), which is then turned back into text.
> and their worth
Tokens are priced based on the model you're using. In general, expect to pay more when you need to use larger models. (Note: Exceptions exist, but treat them as exceptions)
The space as it is now prices tokens based on:
- how many tokens it needs to process (input),
- how many it needs to generate (output), and
- how often you send the same chunks of text that their system already has on file (cached).
> and harnesses
A harness is something that adds in tools & directions for the LLM, so that it can do its tasks. In a dumbed-down semi-visual summary:
---
[Text about what the LLM should be & what they're allowed/not allowed to do]
[Text about available tools & how to call them]
[Your text here]
---
The harness, if its part of a program/app, then looks at what the models generate in response, and does the (file writing/reading, tool calls, web searches, subagents, etc.), based on the formatted output that the LLMs should produce.
> and what exactly all of those things accomplish
Depends on the goal of the user. Some use them as assistants, most use it for coding.
https://openrouter.ai/rankings?utm_source=share&utm_medium=c...
This is insane. He's turning the failure of his AI assistants to do what he wants every single time and to do it in a safe manner into a positive thing. Failure as a positive!
That said, I approach AI with deep skepticism. I try to learn it, including its strengths and weaknesses. But I don't blindly trust it.
Coincidentally, my stock brokerage just yesterday sent an email reminding users that anything their agents do using the user's credentials is the responsibility of the user.
Since failure is such a likely outcome in a startup, "failure as a positive" is a mindset that a startup accelerator must propagate to keep founders going. And when you sell a message for long enough, you start to believe it yourself.
Garry Tan is fully aligned with Elon Musk in going after MacKenzie Scott's charitable giving, making spurious arguments that she's doing it in the wrong way when they are all just running cover for the fact that she's one of the few at that level of wealth being charitable at all these days.
Silly woman, she should give her money away the right way, which I guess is by attempting to buy political power.
https://garryslist.org/posts/mackenzie-scott-s-26-billion-su... - The rant itself.
> This isn’t generosity. It’s negligence at scale.
Is when I realized I was wasting my time reading the article
2020s, techbros: "Here's why it's a huge problem for non-billionaires to have things."
"Garry, this is a text file".
Software? I'm not aware of anything.
> He describes using AI agents as getting help with “almost anything that you do in your day-to-day life. (...) Almost anything you want to do, it knows you, and it’s there to help. And it will actually sort of run ahead and say, ‘What about this?’ Or it’ll help you see around corners.”
Wow, impressive. Both the capabilities and how much he's willing to share.
Is it worth $17k per month? For me, absolutely not. But people spend more on luxurious cars and such, so if you've got money to spare...
> Y Combinator is also using an agent that is “rebuilding the way we operate at YC. 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. It actually has the total context of every founder we’ve ever backed, anyone who’s ever emailed us, like all the applications.”
Holy anthropomorphization. But besides that, what's the goal of doing this? Keeping a handy overview of when and why someone said something?
Because they can. It’s as simple as that. To find meaning, to feel in control, to win. To dominate.
You'd be surprised how important such an overview can be in relationship-based businesses. People spend a lot of time trying to track down a vague memory they have of someone telling them something. CRM platforms are big for a reason, after all.
The only way I can imagine to burn that many tokens is just to run every task fully through the most expensive LLM with no real strategy. Like indexing emails should be a combo of scripts and small summarizing prompts I'd think.
(Figures vary, see e.g. https://www.linkedin.com/posts/christham_4-weeks-on-claude-2... for $13,700/mo.)
If you hit your limits on two Max 20x plans, you could very well be exceeding $300k/yr in token use at current API rates.
git clone https://github.com/garrytan/gstack
cd gstack/
git ls-files | xargs wc -l> he’s suffered from “cyber psychosis” and sometimes barely sleeps
Not cloud costs, though; consumer costs. These agents must run on-device for privacy protections, because anything less will be a global non-starter. That means we need models that excel on phones, tablets, laptops, and desktops; we need home compute modules to run larger, “frontier”-class models for bigger tasks but without guzzling electricity (the Mac Mini’s power envelope is an ideal target); we need smart harnesses that can foist off inference to bigger models elsewhere on your private network of devices as needed without the user prompting for specific thinking modes or model names. All of this has to happen automatically, privately, securely, without leaking data to advertisers or surveillance companies.
It’s hard work, but it’s definitely the future of AI. I would argue the business case is solved with being able to build bespoke stacks of kit and run models through neocloud providers with ZDR policies. It’s the consumer angle that’s hard to crack, because everyone wants a repeat of Web 2.0 for ad-dollars instead of a privacy-centric modern tech stack that forces companies into lower revenue streams.
The future of personal assistants is most likely akin to Muse Spark, and yes I expect Meta is training on all of those interactions and using them to target ads.
I beg to differ.
"People just submitted it. I don't know why. They 'trust me.' Dumb fucks."
- Mark ZuckerbergCounter-point, they can't serve as a knowledgeable discussion partner, programmer, search engine, or your personal psychologist. Not saying that they should.
That's basically gstack
Also...
> Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
It's pretty obvious $200k/yr in tokens is a ludicrously huge spend. Trying to pretend it's not to show how leet you are is weird. But the interesting topic buried here is what value he's creating with such a spend and how it benchmarks against alternatives.
Because that is the value prop of AI. If it doesn't stand up against options when measured clinically and objectively then what's the point of it
Gary is rich. Let's assume now, he spends between $10-20K/month USD on tokens. Whether he gets overwhelming tangible value out of that token spend is immaterial.
What he's getting out of it is a mix of social signaling value and actual perceived productivity. I don't know, but I doubt he has measuring the ROI on the token investment.
Much like someone gets various kinds of value out of a a personal tutor for their child. Or a high powered executive who continues to request budget for a "Chief of Staff" and related personnel in 2026.
The question is whether this is signal, or noise. Will everyone who has the budget aspire to spend money on tokens as a subtle way of projecting value with some tangible benefit? How long of a runway does this have?
Much like a Chief of Staff office, there's more to manage...but you you get some tangible benefit by having people to whom you can [in theory] delegate tasks and trust they will be completed. To some degree none of this matters, value is in the eye of the beholder.
What a time to be alive. Rich people continue to spend money with abandon, just on different things.
Edit: I work at technology company. If I were spending $10-20K/month USD in tokens I would immediately get flagged by Finance and receive a note from my EVPs Chief of Staff/Business Admin. My usage and prompts would get audited and I would be asked to create a report justifying the business impact of the spend.
Does it matter that he's spending this much money on tokens? What's the derived signal? Why isn't he spending more?! Spend on intelligence seems to have no limit, and he's rich after all right? What's the limiting rate of return on token spend for the rich? Does Gary represent an archetype of other rich people who also want to spend money on tokens at that rate? ...if not more?!