Context Window Pricing

Context window pricing is the practical consequence of long-context models: the window is capacity, and using it is a recurring cost. A 200,000-token window does not mean 200,000 free tokens; it means you may spend that much per request if you choose to. The trap is that filling the window is the easiest way to improve quality. Paste the whole document, attach the full history, include every retrieved passage — accuracy goes up, and so does the bill on every subsequent request in that conversation, because the history is resent each turn. A ten-turn conversation with a growing transcript pays for the early turns ten times. Three mitigations are standard. Retrieve instead of pasting, so only relevant passages travel. Summarise the conversation past a threshold rather than resending it verbatim. And use prompt caching where the provider supports it — a long static system prompt or document prefix billed once and reused is often the single biggest saving available. Measure input tokens separately from output. Most teams optimise the output length and discover later that inputs were ninety per cent of the spend.