An AI agent can burn 100× the tokens of a chat turn

An AI agent session can consume 50,000–200,000 tokens, compared to a few hundred to a couple thousand for a chat turn, with multi-agent jobs exceeding one million tokens, according to illustrative ranges from practitioner guides. Gartner's published theme estimates 5–30× a standard query, with coding agents up to 100×, while a Stanford Digital Economy Lab study (arXiv:2604.22750) found agentic coding uses ~1,000× the tokens of code-chat on SWE-bench.

Price · 12 Aug 2026 · Illustrative confidence An agent can burn 100× the tokens of a chat turn A chat turn is hundreds to a couple of thousand tokens. A complex agent session can run 50,000–200,000. Multi-agent jobs go past a million. Download PNG /charts/agent-tokens.png /c/agent-tokens How to read this Illustrative ranges from practitioner guides — not a single official meter. Gartner’s published theme is 5–30× a standard query coding agents up to 100× . Stanford/Brynjolfsson arXiv:2604.22750 found ~1,000× tokens on agentic coding vs code-chat SWE-bench . Sources Practitioner ranges · Gartner theme · Stanford Digital Economy Lab 2026