What will this prompt cost?

how this counts

• It is an estimate, and deliberately a range. An exact count needs the tokenizer the model was trained with — a different vocabulary of a hundred thousand merges per family, megabytes each. Shipping one and quoting it for the whole table would look precise while being wrong for most of it. The same token count is used for every model here for the same reason.
• What it does instead is count structure. Short words are one token; long ones split roughly every four characters; digits break into far smaller pieces than letters; punctuation is usually its own token; and CJK, which has no spaces, runs closer to one token per character.
• The output box is the one that matters. Output costs several times input almost everywhere, and a reasoning model charges for thinking tokens you never read. A long answer from a cheap model can cost more than a short one from an expensive model.
• A cheaper neighbour is not automatically a better buy. The overall column is there so you can see what you would be giving up, and a model with no ranking never appears as a recommendation — only as a neighbour.
• Prices are the vendor's standard pay-as-you-go rate as of 2026-07-25, excluding batch, cached-input, off-peak and priority tiers. Open weights with no first-party API are not priced here: what they cost is your own hardware.