tokens are getting more expensive

The post argues that the usual expectation of cheaper models restoring margins does not hold in AI. Users move quickly to the newest frontier model, and that model tends to stay expensive, while agentic systems and longer-running tasks consume far more tokens than older chat usage.

That combination turns flat-rate pricing into a losing bet. The post says AI companies either need usage-based pricing, very high switching costs, or vertical integration into adjacent infrastructure, because unlimited subscriptions keep getting squeezed by rising token consumption.

Reading notes

#

• Frontier demand shifts to the newest best model, so cheap older models do not solve pricing pressure.
• Token consumption has grown sharply as agents spend longer planning, reading, checking, and rewriting.
• A $20 monthly subscription cannot support a user who can trigger expensive daily runs.
• Claude Code’s unlimited tier is used as an example of how even expensive plans can break under heavy token use.
• Flat-rate pricing creates a prisoner’s dilemma because one company’s subsidies force others to match them.
• The post names three possible exits: usage-based pricing, high switching costs, or vertical integration into hosting and other infrastructure.
• The argument closes by saying that early growth without a path to unit economics can lead directly to failure.