Capabilities include contextWindow and reservedOutputTokenSpace. Reasoning models may reserve additional output tokens. When the meter climbs, summarize or start a new chat.