Summary
Sam Altman once promised that AI would get roughly ten times cheaper each year, yet many developers and businesses are seeing their AI bills climb. This video from Theo of t3.gg digs into why, and argues that intelligence did get cheaper; it just isn’t being measured the right way.
The analysis starts with the headline problem: per-token prices for newer frontier models such as GPT-6 Astra and Fable are higher than those of earlier models like GPT-5. Theo then explains why token price alone is a misleading metric. Agentic workloads generate far more tokens, and factors like caching, reliability, and token efficiency change the real cost of a task. He turns to Artificial Analysis charts that plot cost against intelligence and show the Pareto frontier of the best model at each price point.
The key comparison holds capability constant. DeepSeek V4 Flash, which roughly matches GPT-5 Codex on overall intelligence scores, is about 21 times cheaper on input tokens and 55 times cheaper on output tokens. Artificial Analysis data cited in the video shows the cost of a fixed intelligence tier falling from about 60 cents to 2 cents per task in roughly two years. Viewers get a framework for judging AI costs that goes beyond sticker prices.
📺 Source: Theo – t3․gg · Published October 12, 2026
🏷️ Format: Opinion Editorial
