Prompt Prefill Token Cost Calculator

Prompt prefill token cost calculator

Output tokens are the expensive half of an LLM bill — typically three to five times the price of input tokens. Prefilling the assistant’s response lets you supply the boilerplate yourself instead of paying the model to generate it. This tool quantifies exactly how much you save per request, per day and per month.

How it works

When you prefill, you provide the first part of the assistant’s reply and the model continues from there. You are billed only for what the model generates, not for the prefix you wrote. So the saving per request is the number of output tokens the model no longer has to produce:

saved_output_tokens = tokens the model would have generated for that prefix
saving = (saved_output_tokens / 1,000,000) × output_price

Multiply by your daily request volume to get daily savings, and by thirty for a monthly figure. The savings compound with scale — a tiny per-request gain becomes meaningful across millions of calls.

Practical prefill patterns

Forcing JSON output — instead of asking “Respond in JSON format”, prefill { and the model begins its output from the first key. This is the most common and reliable use case. The savings depend on how much preamble the model would otherwise write before starting the JSON:

// prefill: {"result":
// model continues: [1, 2, 3], "status": "ok"}

Skipping preambles — many models open responses with “Sure, here is the…”, “Certainly, I can help you with…” or similar phrases before the actual content. Prefilling the first word of the real content jumps straight to the answer. Even 20–30 tokens per request adds up to significant cost at millions of calls.

Structured reports — if every response should start with a fixed header (a timestamp, a document title, or a section label), supply it as prefill. The model inherits the structure without generating it.

Partial answers — for chain-of-thought or fill-in-the-middle tasks, a partial sentence can steer the model’s continuation. For example, prefilling “The error is caused by” before asking for a diagnosis can reduce meandering.

Prefill and safety

Some prefill patterns interact with model safety systems. Very directive prefills that push the model into a constrained or contradictory state can produce lower-quality or unexpected completions. Always test your prefill on a representative sample before deploying at scale. Providers may also differ in whether they allow prefills that start mid-word, end with whitespace, or contain only partial JSON structures.

Tips and notes

The biggest wins come from removing predictable boilerplate: opening JSON braces, fixed report headers, “Sure, here is…” preambles, or a known sentence stem. Prefilling does double duty — besides cutting cost it forces the output into the shape you want, which is the most reliable way to get clean JSON out of a model. Keep prefills short and unambiguous; an overly long prefill can box the model into a structure it cannot complete naturally. Always test that the model continues coherently from your prefix, and remember that some providers strip trailing whitespace from the prefill, which can affect formatting.