One billion personal AI agents: the chips, memory, power and money it would take
A first-principles model of a world where every person has an AI agent, with three usage scenarios and every assumption disclosed. Claude Duet was announced October 21, 2026 as a US Pro and Max waitlist research preview. Anthropic confirms a cloud VM and browser, Opus 5.5, and background checks every few hours that share Claude usage limits. VM shape, internal operating cost and independent task performance are not disclosed; Duet contributes no numeric inputs to these estimates.
Last updated . Model pricing is refreshed twice daily.
Overview
A first-principles model of a world where every person has an AI agent, with three usage scenarios and every assumption disclosed. Claude Duet was announced October 21, 2026 as a US Pro and Max waitlist research preview. Anthropic confirms a cloud VM and browser, Opus 5.5, and background checks every few hours that share Claude usage limits. VM shape, internal operating cost and independent task performance are not disclosed; Duet contributes no numeric inputs to these estimates.
- Tokens, GPUs, KV-cache memory and CPU cores per scenario
- Why memory, not compute, is the binding constraint
- Power draw and cost per person per month
- Tools teams can build from this use case
About Tokenomy
Tokenomy is the economic runtime for AI — the system of record for AI model pricing, plus free tools, independent research and runtime rails that meter, enforce and attribute every model call.