Methodology: how AI API prices are verified
Short answer: Every rate on the site is read by hand from one of the 6 official pricing pages listed below and dated in src/prices.json. The numbered steps show how a provider page becomes a row in that file, and the limitations state what the method cannot tell you.
Sources
One official page per provider
Every price on the site traces to one of these 6 pages, each linked below. The “last checked” column is the newest check date among that provider's rows in src/prices.json; a rate is never copied from a reseller, a marketplace or a round-up post.
How to read the table
| Source | URL | Used for | Last checked |
|---|---|---|---|
| Anthropic pricing page | https://docs.anthropic.com/en/docs/about-claude/pricing | Input, cached-input and output list rates for the 13 Anthropic models tracked here (13 verified) | 2026-11-15 |
| DeepSeek pricing page | https://api-docs.deepseek.com/quick_start/pricing | Input, cached-input and output list rates for the 2 DeepSeek models tracked here (2 verified) | 2026-11-15 |
| Google pricing page | https://ai.google.dev/gemini-api/docs/pricing | Input, cached-input and output list rates for the 10 Google models tracked here (10 verified) | 2026-11-15 |
| Mistral pricing page | https://docs.mistral.ai/inference/pricing | Input, cached-input and output list rates for the 6 Mistral models tracked here (6 verified) | 2026-11-15 |
| OpenAI pricing page | https://platform.openai.com/docs/pricing | Input, cached-input and output list rates for the 21 OpenAI models tracked here (21 verified) | 2026-11-18 |
| xAI pricing page | https://docs.x.ai/docs/models | Input, cached-input and output list rates for the 3 xAI models tracked here (3 verified) | 2026-11-15 |
If a percentage or tier on this page disagrees with a provider page, the provider page wins. Send the model name and the URL through the contact form and we will re-check it.
How it works
- Open the provider's official pricing page from the source list above. Reseller pages, model marketplaces and round-up posts are not used as sources. This step is manual.
- Match the exact model name and ID. A similarly named model on the same page is not a match.
- Write the standard, short-context input and output rates — plus the cached-input rate only where the page lists one — into that model's row in
src/prices.json, together withsource(the URL read) andchecked(the date it was read). This step is manual. - Write the facts that are not per-token rates — context window, batch multiplier, long-context tier, cache-write rates — into
src/specs.json, taken from the provider's own model documentation. This step is manual too. - Run
python3 build.py. It reads both files and renders every price table, all 55/models/*pages, the 6 provider pages and the calculator's model list from them, so no page carries a hand-typed price. - The browser calculator in
src/app.jsapplies the same formula ascost()inbuild.py: (uncached input × input rate + cached input × cached rate + output tokens × output rate) ÷ 1,000,000.test/prices.test.mjsasserts that the Python and JavaScript results agree for every verified model. - Copy any scheduled change, promotional rate or intro price into the row's
note, with the provider's date if it gives one; announced changes are also listed on price changes. - If the page shows no readable rate, keep
nullin the row and let the model page say so. Unverified rows stay out of the rankings, the totals and the calculator; nothing is estimated to fill the gap.
Known limitations
- A snapshot, not a history.
src/prices.jsonkeeps one current rate per model plus its check date. When a provider changes a price the row is overwritten, so past rates are not archived here; use your own invoices or the provider's page for a historical record. - Manual and event-driven. There is no automated crawler and no fixed schedule, so a change can appear here after the provider announced it, and a check date is the date a person read the page, not the date the price took effect.
- List rates only. The main calculator and the rankings apply standard, short-context list prices. Long-context tiers and batch multipliers are applied only in the worked examples on model pages, and negotiated, enterprise and reseller prices are not used at all.
What a checked date means
Verified rows currently carry 1 checked 2026-10-27, 51 checked 2026-11-15, 3 checked 2026-11-18. The site-wide date, 2026-11-18, is the newest row date. It is the snapshot date for the price data, not the date every row was read and not the date a provider's price took effect.
| Field | What it records |
|---|---|
Row checked date | The date we opened that model's official pricing page and copied its standard input, cached-input and output rates. |
Site updated date | The newest row date, shown in the footer and used as the pages' dateModified. |
| Not recorded | The effective date of the provider's price, price changes after the check, or the amount on your invoice. |
Cadence: manual and event-driven
Verification is manual and event-driven. There is no automated daily crawl, so a check date means “a person read this page on this date”, not “a job verified it overnight”.
| Trigger | What happens |
|---|---|
| A provider announces a change or publishes a new rate | We re-open the official page, update the affected rows and move their check dates. |
| A site-wide price update | We re-check the verified rows and refresh the site-wide date to the newest row date. |
| A reader reports an error | We open the linked official page and re-check that model; the corrections page records what was fixed, and the contact form is the route. |
| No fixed schedule | We do not promise a daily or weekly check. The per-row date is the only freshness claim. |
What the calculator does and does not apply
| Item | Applied | Where to see it |
|---|---|---|
| Standard input and output rates | Yes | Every model page and the main calculator |
| Cached input % | Yes, when the provider lists a cached rate; otherwise the normal input rate is used | “Cached input %” in the calculator |
| Long-context tiers | Not switched automatically by the main calculator; model pages show the tier and apply it in their worked examples | Per-model “Every price tier” table |
| Batch API discounts | Not applied; the calculator uses standard rates | Batch API guide |
| Promotions and scheduled changes | Not applied to the current rate; listed as notes | Price-change list |
| Taxes, negotiated and enterprise rates, reseller pricing | Not included | — |
The log of announced price changes stays on the price-change list.
Questions
What exactly counts as verification on this site?
A rate is verified when a person opened the provider's own pricing page, read the standard input, cached-input (where published) and output rates for that exact model, wrote them into src/prices.json with the source URL and the read date, and the build rendered the page from that file. A round-up post, a reseller page or another site's table is never a source.
How often are AI API prices checked, and why is there no automated daily crawl?
Manually and event-driven, not on a fixed schedule: a wrong automated number is worse than a slightly older verified one. We re-check when a provider announces a change, when the price file is refreshed (2026-11-18 currently), or when a reader reports an error. Each row's check date is the last manual read, and there is no automated daily crawl.
What happens when a provider changes a price?
The row is updated and its check date moves; if the provider states a future or promotional rate, that sentence is copied into the row note and listed on the price-change log at /price-changes. The database keeps one current price per model, not a history, so your own invoices are the archive.
What does unverified mean?
The official pricing page did not show a readable rate when we checked. The row stores null instead of a number, the model page says so, and the model is excluded from the calculator, the rankings and every total. 0 of the 55 tracked models are unverified in the current snapshot.
Why does a model show no cached price?
Some providers list input and output rates but no cached-input rate. The model is still verified; the calculator charges cached tokens at the normal input rate, and the model page says no cached-input rate is listed. That is different from an unverified price.
Does the calculator apply long-context or batch prices?
The main calculator uses standard short-context list rates. Long-context tiers are shown on each model page and applied in its worked examples, but not switched on automatically by the token calculator. Batch discounts are not applied; the batch guide explains them.
How do I report a wrong price?
Send the model name and the official URL through the contact form. We open the page, re-check the row and move its check date. The correction is reflected on the model page and in the price table.
Can I reuse the price data and the tables?
Yes, the monthly index is published under CC BY 4.0 with a CSV download; cite aiapicostcalc.com and link the page. The site's own pages may be quoted with a link back, and the terms of use at /terms cover high-volume scraping.
Next step → Read the price-change log
Price data last checked 2026-11-18 against official sources · how we verify
Price-change alerts
One short email when a provider changes a price or a promo ends. No spam, unsubscribe any time.