NEW RUNTIME

Topic hub

Agent economics

How AI work is priced, measured, routed, and justified by completed task value rather than raw token volume.

Short answer

  • Agent economics moves measurement from token use to useful completed work.
  • Cost only makes sense beside success rate, review time, retry loops, latency, and human attention.
  • The mature metric is cost per accepted outcome, not cost per generation.

Pattern memory

Current hypotheses

1 patterns
  1. high

    Agent economics moves to completed work

    The economically meaningful unit for agent systems is becoming cost per verified completed task rather than cost per token or model call.

Field notes

Longer analysis

1 notes

  1. Gemini 3.6 Flash Moves the Agent Race Toward Cost per Task

    Google's Gemini 3.6 Flash release frames the model race around token efficiency, built-in computer use, and specialized cyber agents rather than raw chat intelligence alone.

Raw signals

Recent observations

6 signals

  1. A pricier model can be cheaper per completed task

    Cognition reports that Fable 5 completed coding work with fewer steps and output tokens than its previous lead model.

  2. Claude Code separates model choice from effort

    Anthropic exposes model selection and effort level as different controls for capability, token use, latency, and persistence.

  3. Databricks benchmarks coding agents on its own codebase

    Databricks evaluates agents on fresh internal pull-request tasks and measures success alongside runtime, tokens, and cost.

  4. An advisor model can guide a cheaper executor

    The advisor-tool pattern lets a fast executor request bounded analysis from a stronger model while keeping control of the task loop.

  5. Sonnet moves agent capability down the price curve

    Anthropic positions Claude Sonnet 5 for planning, terminal work, browser use, and multi-step agent tasks at a lower tier.

  6. AI spend should be measured per successful task

    Ramp argues for allocating AI cost by use case, owner, completed outcome, failures, retries, review effort, and latency.