Y Combinator’s Garry Tan Spends Up to $200K a Year on AI Agent Tokens
TL;DR
- Y Combinator president Garry Tan says his intensive use of AI coding agents could cost as much as $200,000 a year in model tokens.
- Tan’s comments—partly serious and partly framed as jokes about “cyber psychosis”—highlight the extreme habits emerging around always-on AI tools.
- His experience points to a broader shift: advanced AI experimentation is becoming less constrained by software access and more constrained by the cost of sustained, high-volume usage.
The New Price of AI Experimentation
For many people, AI assistants are a modest monthly expense. For Y Combinator president Garry Tan, they may represent something closer to a six-figure operating cost.
Tan has described an unusually intense workflow built around AI agents—software systems that can reason through tasks, write and modify code, operate across files, and continue working with comparatively little supervision. He has suggested that, at his rate of use, annual spending on model tokens could reach as much as $200,000.
The figure is striking not only because of its size, but because it offers a glimpse of where AI experimentation may be heading for heavy users. The central cost of using advanced models is increasingly shifting from access to capability: the more tasks an agent performs, the more context it consumes, and the more frequently it is asked to iterate, the larger the bill becomes.
Tan’s AI-First Workstyle
Tan has portrayed AI agents as deeply integrated into his daily work rather than as occasional productivity tools. Instead of using a chatbot to answer isolated questions, he appears to treat agents as persistent collaborators capable of handling multiple technical and operational tasks.
That style of use can involve asking agents to inspect codebases, propose changes, run tests, debug failures, and revise their work repeatedly. Each action may consume additional tokens, particularly when the system needs to retain a large amount of context or review the results of earlier steps.
For an executive overseeing a startup accelerator, the potential applications are broad. Agents can help analyze companies, prepare documents, draft communications, build internal tools, and explore new product ideas. They can also act as rapid research assistants, software developers, and brainstorming partners.
The trade-off is that an agent working continuously can generate far more usage than a conventional chat session. A single request may trigger a long chain of model calls, with each call adding to the total.
How a $200,000 Bill Could Accumulate
Tan’s estimate should be understood as a description of possible usage rather than a standard price paid by most AI customers. The total depends on the models used, the size of the prompts and responses, the amount of context retained, and how many agents are running at once.
Token-based pricing generally counts the text processed by a model, including both the user’s instructions and the material supplied as context. In agentic systems, that context can include source code, tool outputs, previous conversations, logs, and task instructions. Repeatedly sending that information back to a model can make complex workflows expensive.
High-end models are also often used for the tasks where reliability matters most: writing production code, diagnosing difficult bugs, planning multi-step changes, or checking the output of another model. If an agent is allowed to retry, review, and refine its work, the token total can rise quickly.
A six-figure annual estimate would amount to roughly $16,700 per month, or more than $500 per day averaged across a year. That is far beyond the typical consumer subscription, but it is not necessarily implausible for someone operating multiple agents at high intensity and using premium models for extended periods.
The number also illustrates an important distinction between software pricing and software consumption. A flat subscription may make AI appear inexpensive, while intensive users can incur substantial additional costs through usage-based plans, premium model access, application programming interfaces, or custom infrastructure.
The “Cyber Psychosis” Joke
Tan has also used the phrase “cyber psychosis” while discussing the effects of spending extreme amounts of time with AI systems. The expression was presented humorously, but it reflects a recognizable concern among AI enthusiasts: the possibility that constant interaction with agents can distort a person’s routines, attention, and sense of boundaries.
The joke lands because agentic software encourages a different relationship with computers. Traditional applications wait for users to click buttons or enter commands. AI agents can appear more conversational and proactive, inviting users to delegate increasingly large portions of their work.
That can be exhilarating. It can also make it harder to disengage. When an agent is capable of continuing to investigate a problem, generating new ideas, or implementing another improvement, there is always a reason to keep going.
Tan’s comments should not be treated as a medical diagnosis or evidence that intensive AI use causes a defined psychological condition. In context, “cyber psychosis” is a colorful description of the strange, immersive, and sometimes obsessive experience of working alongside increasingly capable software.
Sleepless Work and the Productivity Race
Tan has additionally discussed working extremely long hours and sleeping very little. Those remarks fit a broader culture in technology in which founders and executives describe AI as compressing the time required to build, launch, and evaluate products.
AI agents can amplify that culture by making more work possible at any hour. A person can ask an agent to continue coding overnight, review output in the morning, and then launch another round of tasks. The technology effectively expands the number of parallel workstreams a small team can manage.
But greater output does not automatically mean better judgment. Human oversight remains necessary for deciding what should be built, checking whether an implementation is secure, and distinguishing a useful result from a confident but flawed one. Sleep deprivation can make those decisions more difficult, particularly when an AI system produces a steady stream of plausible suggestions.
The tension is central to the current AI productivity narrative. Agents may reduce the time needed to complete individual tasks, yet they can also create more tasks to review, more experiments to run, and more outputs to monitor.
What Tan’s Experience Says About AI Costs
Tan’s spending estimate is notable because it comes from a prominent technology investor and executive rather than from an ordinary consumer testing an AI app. His experience suggests that the economics of advanced AI tools are becoming increasingly uneven.
For casual users, a monthly subscription may be sufficient. For developers, researchers, founders, and organizations that run agents continuously, costs can scale with ambition. The more context an agent handles and the more often it acts, the more expensive the workflow can become.
This creates a new form of infrastructure spending. In the past, a small software team might have budgeted for cloud servers, databases, developer tools, and contractors. Increasingly, it may also need to budget for model inference—the computing required to generate AI outputs.
The cost may eventually fall as models become more efficient, hardware improves, and providers introduce better caching and lower-cost versions. Companies are also developing systems that route simple tasks to cheaper models while reserving expensive models for difficult decisions.
For now, however, the most capable workflows can carry a substantial premium. The technical ability to run an agent is becoming widely available, but running one intensively remains an economic choice.
A Signal for Startups and Investors
Y Combinator’s startup network gives Tan’s comments additional significance. Founders are being encouraged to build companies that rely heavily on AI from their earliest stages, sometimes with very small teams and large amounts of automated work.
That model can be attractive because agents may let startups prototype products, conduct research, write code, and support customers before they have the resources to hire large departments. But token costs can complicate the assumption that AI automatically makes software cheaper.
A company that gives every employee access to powerful agents may see productivity gains alongside rapidly increasing usage bills. If agents are embedded in customer-facing products, the business must also account for model costs each time a customer triggers an interaction.
The result is a new discipline for startup planning: measuring not only how much time an AI system saves, but also how much it costs to produce that saving. A workflow that looks efficient in a demonstration may be uneconomical when it runs thousands of times.
The Human Limit
The most revealing aspect of Tan’s account may not be the $200,000 estimate itself. It may be the way his use of AI exposes the limits of human attention.
Agents can work faster, operate continuously, and handle several processes at once. Humans still have to decide which outputs matter, approve consequential actions, and maintain a coherent understanding of the systems they are building.
That imbalance can produce a strange form of abundance: more code, more drafts, more experiments, and more possible directions than one person can comfortably evaluate. The challenge shifts from generating ideas to selecting, verifying, and stopping.
Tan’s “cyber psychosis” comments capture that tension in deliberately exaggerated language. AI can feel like an extraordinary force multiplier, but it can also encourage users to remain permanently engaged with a stream of machine-generated work.
The Next Phase of AI Adoption
Tan’s reported spending offers an early look at a future in which the most enthusiastic AI users treat models less like applications and more like an always-available workforce. That future could bring major gains in software development and entrepreneurship, but it will also expose new questions about cost, oversight, labor, and personal sustainability.
For now, the $200,000 figure is best viewed as a warning about the upper end of AI consumption—not a forecast for ordinary users. Yet it demonstrates how quickly expenses can grow when advanced models are used continuously, with large amounts of context and multiple rounds of automated reasoning.
As AI agents become more capable, the definition of “using AI” will become less meaningful without a measure of intensity. The important questions will be how many agents are running, which models they use, how often they retry, how much context they process, and whether the resulting work justifies the bill.
Tan’s experience suggests that the next AI cost curve may be shaped not only by model prices, but by how difficult it becomes for people to stop experimenting.
Get All The Latest Updates Delivered Straight To Your Inbox For Free!