This operator's other projects don't ship unsourced numbers, and neither does this one. Here's what every term means and exactly where every figure comes from.
Longer, worked guides built from the same dated price data as the calculator. Every number on them is recomputed from our dataset when prices change.
Every major provider charges input and output tokens at different rates — output is almost always more expensive, between 2 and about 8 times the input rate (5 times is the most common ratio among the models tracked here), because generating text costs more compute than reading it. A "token" is roughly ¾ of a word in English (so ~4 characters); code, non-English text, and text with a lot of punctuation tends to tokenize less efficiently than plain English prose. Providers publish exact tokenizers (e.g. OpenAI’s tiktoken) if you need an exact count rather than an estimate — the rough-guide ratio used on the Calculator page is an approximation, not a substitute for actually counting tokens for a cost-critical workload.
Several providers (Anthropic, Google, xAI) offer a much cheaper rate for input tokens that repeat across requests — typically a long system prompt, a shared document, or conversation history you resend each turn. This can cut effective input cost by 70-90% for workloads with a lot of repeated context (a coding assistant re-sending the same file, a chatbot re-sending conversation history). Not every model publishes a cached rate; where one isn’t listed, this site’s calculator falls back to the full input rate rather than guessing a discount.
LLM API pricing changes more often than almost anything else this operator has built a calculator for — weeks, not years, between changes. Every figure on this site carries a lastVerified date and one of two confidence levels:
This site will re-check every price roughly monthly. If you spot a stale or wrong figure, the fastest fix is checking the linked source yourself — this is a reference tool, not a live feed.
No fine-tuning/training cost estimates (pricing structures vary too much by provider and use case to generalize safely), no image/audio/video generation pricing (a different cost model entirely — token counting doesn’t apply), and no live currency conversion (all prices are USD, matching how every provider bills — a static conversion rate would go stale faster than the prices themselves and imply false precision). These are scope decisions, not oversights.