Methodology
Everything on this site rests on one dataset and fourteen sets of arithmetic. This page documents both, including the parts that are approximate — because a cost tool that hides its assumptions is worse than no cost tool at all.
How prices get into the index
Each entry in the index is read from the vendor's own published pricing page. Not from a comparison site, not from a press release, and not from a model's recollection of what things used to cost.
An automated job runs weekly. For every tracked item it fetches the vendor's pricing page, extracts the current figure, and compares it against what is on record. Three things can happen:
- The price matches. The entry's verification date is updated and nothing else changes.
- The price has moved. The new figure replaces the old one, and the change is written to the price change log with both values, the date, and the source URL.
- The page could not be read or parsed. The old figure is kept and flagged as stale. Nothing is guessed. An entry that cannot be verified for three consecutive weeks is marked as unverified on the price index rather than quietly presented as current.
Prices are recorded in US dollars, excluding tax, at published list rates. Volume commitments, enterprise agreements, startup credits and regional variations are not modelled — if you have negotiated a discount, your real cost is lower than anything shown here.
Known limitations of the index
JavaScript-rendered pricing tables. Several vendors publish prices that only exist after a page's scripts run. Where a figure cannot be read reliably, the item is either sourced from a second official page — an API documentation page, or a cloud marketplace listing for the same model — or left out entirely. It is never estimated.
Multi-GPU node pricing. Some GPU vendors quote per-node rather than per-GPU. Where that is the case the per-GPU figure shown is the node price divided by the GPU count, and the entry says so.
Tiered and conditional pricing. Rates that vary by context length, by region, or by monthly volume are recorded at their standard tier. Where a vendor's cheapest tier is materially different, that is noted on the item rather than blended into an average.
Model naming. Vendors rename and re-version models frequently. The index uses the vendor's current public name. When a model is renamed, the change log records it as a rename rather than a price change.
What the calculators assume
Every calculator runs in your browser using the same price index. None of them transmit your inputs anywhere — the pages are static HTML with no back end, so there is no server that could receive them.
Where a calculation requires a conversion, these are the constants used throughout:
| Conversion | Value | Why |
|---|---|---|
| Words to tokens | ×1.33 | Typical for English prose across common byte-pair tokenisers |
| Characters to tokens | ÷3.9 | Blended with the word estimate; adjusted for punctuation density |
| Characters per word | 5.7 | Standard English average including the trailing space |
| Spoken words per minute | 150 | Broadcast and audiobook narration average |
| Hours per month | 730 | 8,760 ÷ 12, so monthly GPU figures are comparable across month lengths |
| Vector index overhead | ×1.6 | Raw vector bytes plus typical HNSW graph overhead |
Token estimates are approximations. Every vendor uses a different tokeniser, and none of them run in a browser without shipping a large vocabulary file. The blended heuristic used here lands within roughly 5–10% for ordinary English prose and degrades for code, markup and non-Latin scripts — which is why the token estimator reports a characters-per-token diagnostic rather than just a number, so you can see when you are in the accurate case and when you are not.
What the calculators deliberately do not model
Quality. Not one of these tools knows whether a cheaper model will do your job well enough, and none of them pretend to. Cost analysis tells you what is worth evaluating. Evaluation tells you what to ship. Any tool that claims to collapse those two steps into one is selling something.
They also exclude the surrounding engineering cost — building, testing, monitoring and maintaining the feature — except where a field explicitly asks for it. The guide to the costs that never appear on a price card covers what falls outside the arithmetic.
How the calculations are checked
Each calculator's arithmetic is verified against hand-computed values before it ships, and re-verified whenever the underlying formula changes. Where a calculator implements something with a derivation worth stating — the break-even hit rate for prompt caching, the break-even utilisation for self-hosting — the derivation is described in that page's explainer rather than left implicit.
Corrections
Errors get fixed, dated, and logged. If a figure on this site is wrong, the vendor's own page is the authority and we would like to hear about it: hello@themagicofai.com.
Prices change without notice and this index is a snapshot, not a contract. Before committing budget, confirm against the vendor's own pricing page — every entry in the index links straight to it.