By J. Wren · Last reviewed 2026-10-21
Longer write-ups that go past what a single calculator can show — measured numbers, worked examples, and a look at what actually changed rather than a restatement of what a rule of thumb says should happen. Each one starts from a real question that came up while building or maintaining this site, not from a keyword list.
What actually moved, using this site's own change log rather than a vendor's summary of itself.
The rule of thumb tested against real tokenizer output on English, Chinese, code and JSON, with the size of the error in each case.
Embedding and generation costed out for one concrete pipeline, end to end, with the line items that are easy to forget when estimating from memory.
The same fifty text samples counted with Gemini's countTokens endpoint and two OpenAI tokenizers, then priced, with the script to re-run it.
One request that fills the window, priced at full, half and a tenth on 13 models, including the higher rate four of them charge on long prompts.
These are not tool pages: there is nothing to paste in and no result to compute. They exist because some questions are answered better by a number someone actually measured once than by another calculator, and because a pricing site that never says anything about the numbers it shows is not saying much at all.
Looking for a calculator instead? Start on the home page.