Token prices fell 8割 but AI bills rose, Postman writer says

The article builds on the two earlier installments in this series, which examined API keys and access tokens — the credentials agents use to reach outside services — and the risks of handing agents a human's keys. Kusanagi frames credential management as only the entry point to running agents, with cost as the equally overlooked second problem that shows up later.

The mismatch between unit price and total bill has drawn investor attention too. Andreessen Horowitz has analyzed that the token price for the same capability keeps falling at roughly a tenth per year, and the piece notes that GPT-3-class performance dropped from about $60 per million tokens in 2021 to about $0.06 in late 2024 — a nearly thousandfold decline in three years.

The piece also connects tool design directly to cost. Wrapping existing REST APIs one-to-one into MCP tools is easy and works at first, but mechanically generating a server from an OpenAPI spec turns 200 endpoints into 200 tools, inflating context and causing selection errors. A cited study found that narrowing tools to only what is needed raised tool-selection accuracy more than threefold and cut prompt tokens by more than half. Whether enterprises can bend their spending curve downward therefore appears to hinge less on model vendors' price lists than on how carefully they curate the tools and context they hand to agents.