TokenScale is built to verify every price against the provider's official pricing page every night. This page explains what "verified" means here — and, just as importantly, how we flag any day that was carried forward instead of freshly checked, so you always know which is which.
Verified means a human-readable price was read from the provider's official pricing page that night and matched against what TokenScale shows. It does not mean scraped from a third-party aggregator, and it does not mean estimated.
Verified means a human-readable price was read from the provider's official pricing page, matched against what we show, and the provider stamped with that date. The system is designed to do this every night. When a night's check can't complete (a page that won't render without scripts, a provider outage, our own pipeline stalling) we don't guess and we don't pretend: that day is carried forward from the last verified price, drawn in grey, with the "verified" date left at the last real check. When two official sources disagree (it happens more than you'd think), the conflict is noted rather than silently picking one.
A grey day stays grey. This page used to say that once the next verification landed, grey days turned solid. That sentence contradicted the append-only rule above it, and on 4 August 2026 we removed it and fixed the record to match. A successful check on Tuesday confirms Tuesday; it cannot reach back and confirm a Monday we missed, because between the two a price could have moved and moved back without us seeing it. So carried-forward days are now flagged permanently in the data and drawn grey forever. The gap is part of the record. Every missed night, with the reason, is listed on the reliability page.
TokenScale expresses sizes in content you recognise. The word counts behind the landmarks are real: The Hobbit is 95,356 words, verified against the Project Gutenberg text. Tokens are estimated from words at roughly 1.33 tokens per word, the ratio that holds for typical English prose. Different tokenizers vary a little either side of that; the price differences between providers are far larger than the tokenizer differences.
Limited-time promotional rates are shown as an overlay on the live price, clearly labelled. They never enter the price history, so the historical record reflects standard published rates only.
Some providers publish different standard on-demand prices at different times of day. TokenScale records the highest published standard rate as the comparison figure, labels it as a peak rate and shows the exact lower schedule beside it. This gives a conservative planning number without pretending the peak applies all day. We do not average the windows because a provider has not published that average and each customer's traffic pattern is different.
For DeepSeek, peak hours are 01:00–04:00 and 06:00–10:00 UTC. All other hours are off-peak and cost 50% less.
Corrections are published, not buried. The first one set the tone: on launch week the headline figure said writing The Hobbit cost $0.04, which was the input cost only. The correct total was $0.06. The fix, and why it happened, went straight into the changelog. That distinction between input and output pricing is exactly what this site exists to make visible.
If a number here doesn't match the provider's page, we want to know within the day.
Spotted an error? Use the "Report a pricing error" link in the calculator's footer, or email projectsbilton@gmail.com. Confirmed same-model corrections are applied and logged; anything ambiguous is flagged on the site rather than guessed at.
Last updated: 18 August 2026. If the process changes, this page changes with it.