Output Length Cost Impact

Shows what shortening the answer is worth per month, and what 100 extra output tokens on every request cost you.

Requests per month

Input tokens per requestHeld constant here — this tool is about the answer

Output tokens per request now

Output tokens you are aiming forAfter a length cap, a terser format or a shorter instruction

Input rateUS dollars per 1,000,000 tokens. Example figure — replace it with yours.

$/1M

Output rateUS dollars per 1,000,000 tokens. Example figure — replace it with yours.

$/1M

Saved per month by shortening the answer

$1,350

Output is 70.0% of this workload's bill today, and every 100 output tokens you add to every request costs $300.00 a month.

Input cost per request: $0.0045
Output cost per request now: $0.01
Output cost per request at target: $0.0037
Total per request now → target: $0.01 → $0.0083
Output share of the bill now: 70.0%
Output tokens removed per request: 450
Monthly bill now → target: $3,000 → $1,650
Annual saving: $16,200
Cost of +100 output tokens on every request: $300.00 / month
One output token costs as much as: 5.0 input tokens

What this assumes: input length is held constant, so this measures the output lever alone. Reasoning tokens, where a model bills for thinking you never see, count as output — include them in the output figure or the number here will be low. A month is 30 days. The rates are examples; the relationship the tool is really showing is the ratio between them, which survives a price change.

The lever nobody pulls

Output is the expensive half of an inference bill and the half almost nobody tunes. There are two reasons for that. The obvious one is price: output tokens are usually billed at a multiple of input tokens, because generating them is sequential work on a GPU while reading the prompt is parallel. The second is that output length is the term you control most directly — you can ask for less, cap it, or change the format — and precisely because it is easy, it gets left at whatever the model happened to produce.

The useful line in the breakdown is the last but one: what 100 extra output tokens on every request cost per month. That converts a wording decision into money. "Explain your reasoning" is not a free instruction; neither is a response schema with verbose field names, or a system prompt that invites a preamble. On a workload at any scale, a habit of two polite sentences before the answer is a line item.

What this does not capture: shortening an answer can cost you accuracy on tasks where the model genuinely benefits from working through the problem, and a truncated answer that has to be regenerated costs twice. A hard max-tokens cap is not a saving if it produces cut-off JSON that fails validation and gets retried — the retry pays for both. Shorten by asking for less, not only by capping; use the cap as a guard against the tail, and check what share of requests hit it.

Output Length Cost Impact

The lever nobody pulls

The reasoning behind this

Related tools