AI token calculator methodology

1. A maintained local tokenizer is labeled Exact tokenizer.
2. A family-specific local heuristic is labeled Model-family estimate.
3. A language-aware general heuristic is labeled Generic estimate.
4. A count entered by the user is labeled User-supplied token count.

Pasted content is processed locally and is never included in analytics events, URLs or database writes.

Each billable category is calculated independently at full precision: standard input, cache read, cache write, output, reasoning, tools and supported modality units. Batch percentages blend real-time and documented batch rates. When a documented long-context threshold changes all token prices, the whole request receives that tier.

Pricing records have effective windows. Old records remain in history; future or promotional records do not silently replace current production values. A current model requires an official source, verification date and pricing note.

First-party provider documentation is authoritative. A page hash can flag an official-page change, but it never automatically publishes new prices. A reviewer must confirm the meaning, update the versioned record and record the review date. Report suspected errors to hello@piskill.online.

Estimates do not include taxes, region-specific billing, private enterprise discounts, rate limits or undocumented promotions. Provider invoices are authoritative. Model suitability is filtered by recorded technical facts and price—not fabricated quality scores.