YOU

Tokens

Model-inference tokens: units of model input and output used for context limits and cost accounting when running harnesses, workers, or inference.

What It Is

A model-inference token is a unit of model input or output used for context and cost accounting when a harness, worker, or inference step runs. Prompt text, tool results, and model replies all consume these tokens. The model's context window is the limit on how many of them fit in one turn. Model-inference tokens are not factory work tokens that move submitted work through places.

Why It Matters

Token counts drive how much context a run can keep, how expensive a turn is, and when compaction becomes necessary. Raising thinking or tool use can increase token spend; long loops fill the context window until later turns no longer fit. Operators who watch cost, latency, and context pressure are watching model-inference tokens.

Simple Example

You start a long coding goal under a harness. Early turns send a short prompt and get a short reply, so token use stays low. After many tool calls and retries, the retained transcript approaches the model's context window. Compaction shrinks older detail so the next turn still fits—because model-inference tokens, not factory work tokens, filled the window.

Common Confusions

Model-inference tokens are not factory or work tokens. Factory work tokens are the Petri / Colored Petri Net (CPN) units of submitted work that occupy places as workstations fire; that meaning lives in program documentation such as Petri, configuration, workstations, and submitting work. Design-system or UI tokens are also a different idea and are out of scope here. Compaction shrinks retained context measured in model-inference tokens; thinking is deliberative reasoning that may spend more of them.

Tags

On this page

What It IsWhy It MattersSimple ExampleCommon ConfusionsTags