LLM API Cost Calculator
What a month of calls actually costs, including the parts people forget.
The formulas
- Input cost =
requests × input tokens ÷ 1,000,000 × input rate - Output cost =
requests × output tokens ÷ 1,000,000 × output rate - Cached portion =
input cost × cache share × (1 − discount)plus the uncached remainder at full rate - Monthly total =
input cost + output cost
Output tokens usually dominate
Output is priced several times higher than input because generation costs far more compute than reading. A chatbot with a 1,500-token prompt and a 500-token answer looks input-heavy, but at typical rates of 3 dollars per million input and 15 per million output, the input costs 0.0045 dollars per call while the output costs 0.0075 — the answer is more expensive than the question despite being a third the size. Optimisation efforts aimed only at shortening prompts miss the larger line item. Capping max_tokens is often the single cheapest change available.
Where production costs actually come from
- Retrieval dumping. Passing ten retrieved documents when two would answer the question multiplies input cost by five with no quality gain.
- Silent retries. A 5% retry rate is a 5% surcharge that nobody budgets for. Log it.
- Conversational history. Re-sending the full transcript on every turn makes cost grow quadratically with conversation length. Summarise older turns.
- Unbounded generation. Without max_tokens, a single confused response can emit thousands of tokens.
- Development and evaluation traffic. Test suites and evals frequently account for a meaningful share of a monthly bill before launch.
Prompt caching changes the maths
Most major providers discount repeated prompt prefixes heavily — often 50% to 90% — when the same prefix is sent within a short window. That makes the structure of your prompt a cost decision, not just a quality one: put the stable part (system instructions, few-shot examples, reference material) first and the variable part (the user's actual question) last, so the stable prefix can be reused across requests. Done well, caching cuts input cost dramatically on high-volume workloads. Done badly — variable content at the front — you pay full price every time.
Comparing models fairly
The cheaper model is not always cheaper per task. A model that costs five times more per token but needs one attempt instead of three retries can win outright, and a weaker model that requires more post-processing shifts cost elsewhere. Measure cost per successful outcome, not cost per token. Consider also whether the cheaper model handles your longest prompts — a small context window forces chunking, which adds complexity and may erase the saving. Use the token counter to size real prompts, and the context window calculator to check they fit.
Reading the numbers
Prices change frequently and vary by provider, region and committed-use discounts, so treat the rates you enter as the source of truth rather than any default shown here. Volume commitments and batch APIs typically offer substantial discounts for work that does not need to be synchronous — if half your traffic is offline processing, batch pricing can halve that portion of the bill at no quality cost.