Using a large language model feels like a slot machine. You ask a question, the model returns a response full of confidence and specificity, and if it works the payoff is instant. When it fails—especially in coding tasks—your frustration doubles, yet the unpredictability keeps users hooked for developers across industries, it's a daily gamble today.

The economics mirror the game: every input token costs money, and output tokens cost five times more. Anthropic’s Opus 4.6 charges $5 per million input tokens and $25 per million outputs, while OpenAI’s GPT‑4.5 runs at $2.50 and $15. A million tokens equals roughly the length of the Harry Potter series, so usage can snowball for large.

Subscriptions hide the true cost. A Claude user pays $20 a month for a tier that includes code generation and integrations, yet the plan’s limits depend on conversation length, model choice, and feature use—details buried in fine print. Users often discover their monthly cap is reached before they realize how many tokens they’ve consumed and the total expense remains unclear.

The takeaway is clear: running generative AI at scale costs more than advertised, and the pay‑per‑token model forces developers to navigate uncertain output lengths. If the industry continues to frame usage like gambling, it will struggle to justify ROI for businesses that demand predictable, cost‑effective tooling in the long term today.