Why that $132M agent spend figure is wrong: caching and falling token prices

Token counts that include cached tokens can overstate agent spend by an order of magnitude or more. The post gives a worked cost correction and a trend (cost at fixed intelligence falling fast) for budgeting agent usage.

• Theo says his own 33.5B tokens of usage came to $0.56 per million, under a tenth of the $8 per million estimate behind the viral figure, so the real API cost is more likely $100k to $500k a month than $11m.
• Even using expensive models like Opus and Astra, he estimates 1.5T tokens would have cost him about $800k, and cheaper models with more cache hits could bring it as low as $100k.
• He notes Artificial Analysis previously saw cost at a given intelligence level fall about 10x a year, and projects that if the faster trend holds, a $100k/month workload could cost $3k/month within months and about $100/month in a year.
• Theo argues the spend should be judged as a mix of product and engineering, research, forward-deployed work and marketing, not as dollars in for software out, and calls it worth a million or two a month.

• AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
• token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

• postT3 Code ships in-thread visualizations for agents
• postCutting GitHub rate-limit use 75% in an agent PR watcher
• postAudit your agent history for reinvented PR-babysitting scripts
• post3D demos are a benchmark of their own, not a proxy for coding skill

• videoEx-NVIDIA Engineer: Why AI Is About to Get 1000x CheaperInvest Like the Best
• articleWhy the cheap model sometimes costs more than the expensive oneWRITER
• articleManaging a token budget? Think efficiency, not intelligenceWriter Team

Checking sign-in…

Loading comments…