Methodology

OpenAI publishes its tokenizer, so counts for GPT-4o and GPT-5 family models are exact. They are produced by running the real o200k_base byte-pair encoder in your browser.

Anthropic and Google do not ship a public offline tokenizer. Counts for their models are estimates, derived from the o200k count and a family factor. Every place a number appears, the interface says which kind it is.

The factors are deliberately conservative. Claude models up to Sonnet 4.6 and all Gemini models are treated as comparable to o200k for English text — an assumption, not a published figure. The one hard number is Anthropic’s own documentation, which states that the tokenizer introduced with Claude Opus 4.7 produces approximately 30% more tokens for the same text than the previous Claude tokenizer.

38 models, transcribed by hand from published pricing pages and dated in the source. Last verified 2026-08-10.

They are not scraped at runtime. Scraping would add a server, a failure mode and a legal question in exchange for freshness measured in weeks, on data that changes monthly. A dated file that a person updates is the honest trade — and it means the whole application can be static.

Figures are first-party API, default (global) routing, USD. Regional endpoints, data residency multipliers, enterprise discounts and marketplace billing are not modelled.

A cache write costs W per token and each read costs R, against paying the base rate B every time. Over one write plus k reads, caching costs W + kR and not caching costs (k+1)B, so caching wins when k > (W − B) / (B − R).

With a short-lived write at 1.25x the base rate and reads at a tenth, that gives k > 0.28 — one read. With an extended write at 2x it gives k > 1.11 — two reads. Those match the numbers Anthropic publishes, which is the check that the derivation is right; it is asserted in the test suite.

The static prefix is taken as everything above the first per-request value in your prompt, because a cache prefix ends at the first byte that changes.

The headline “safe rewrite saves” number is exact: the rewritten prompt is re-counted from scratch and the difference priced directly, rather than added up from separate estimates.

The per-finding figures are attributions, and they can overlap. Caching examples and deleting examples both save the same tokens, so adding every finding together would overstate the total. That is why the interface reports the rewrite saving and the single largest structural opportunity as two separate numbers rather than one inflated one.

No step calls a model. Every one of the 26 checks is deterministic string and arithmetic work, which is why the tool is free, instant, and never sends your prompt anywhere.

It doesn’t judge quality. It can price what an example costs; whether removing it hurts your accuracy is a question for your evals. Deduplication is off by default for that reason, and the deeper structural findings are advisory rather than automatic.

It has no visibility into your traffic. Every monthly figure comes from the request volume you enter, multiplied out — change that number and every dollar figure changes with it.