How are tokens counted?

Prompt and completion tokens come from OpenRouter, then we add them up for the run.

Tokens for a run are what OpenRouter reports, not a local word count.

Per LLM call

Each model call is streamed with usage enabled. When the stream ends, OpenRouter sends prompt_tokens and completion_tokens.

• Prompt tokens — system prompt, history, images, and tool results sent into that call
• Completion tokens — text and tool-call JSON the model wrote

Per run

A run can make several LLM calls (first reply, then after tools). Galaxy adds those usages together and stores them on the run and the assistant message.

The footer shows prompt + completion as N tokens.

Not in the token count

• Magica tools — those use application credits
• Failed OpenRouter retries — only the attempt that returns counts
• A guessed local tokenizer — if OpenRouter omits usage, that call is 0