Definition
Token burn is the amount or rate of tokens consumed by model activity over a stated boundary, such as one request, one accepted task, one agent run, one user-day, or one billing period. It answers how much inference volume the system used, not whether that use was wise.
A useful report separates input, output, reasoning, cache reads, and cache writes because providers expose and price those categories differently. It includes retries, failed runs, subagents, scheduled work, and background activity when they fall inside the chosen boundary. It also names the model, provider, harness, and time window so later comparisons are meaningful.
Read the burn with an outcome
Raw token totals encourage bad conclusions. A higher-capability model may burn fewer tokens by solving a task in one attempt. A cached prompt may cost less while the model still processes a large prefix. A complex investigation may justify heavy burn and create reusable learning. Compare consumption with accepted outcomes, retained learning, latency, human attention, and durability.
Distinguish it from nearby terms
Token budget is the planned allowance. Token burn is observed consumption. Burn rate adds a time denominator. Token maxing is a strategy or incentive to increase use. In cryptocurrency, token burning means permanently removing assets from circulation; that unrelated sense should be made explicit when audiences overlap.
Check your understanding
Workflow A burns 500,000 tokens per day and Workflow B burns 50,000. Which is more efficient? You cannot tell until you know what work each attempted, how many results were accepted, which costs were cached or hidden, and what later failed.