How AICostBudget Collects and Verifies AI API Prices

Quick answer

What we collect

Each published pricing record can include input, cached input, cache write, batch input, batch output, and output prices. Context window and release date are included when an official source publishes them.

The current published data contains these conditional charge types:

• Input tokens
• Cached input tokens
• Cache reads
• Cache writes
• Cache writes with a five-minute lifetime
• Cache writes with a one-hour lifetime
• Cache writes with a twelve-hour lifetime
• Output tokens
• Storage
• Document pages
• Session duration

Official sources and source hierarchy

1. Official provider pricing page
2. Official provider documentation
3. Official provider API
4. Official provider announcement

Third-party calculators, SEO blogs, social posts, competitor comparison pages, and unattributed screenshots are not accepted as final price authorities. Every displayed price links to its official source in a new tab with nofollow noopener.

What “Last verified” means

“Last verified” is the date a value was checked against an official provider source. “Last checked” records a review that did not establish a fully verified value.

Change-log dates

The published dataset was last verified on Oct 21, 2026. Public record verification dates run from Jul 7, 2026 through Oct 21, 2026.

• 56 records were verified within 30 days of the dataset date.
• 76 records were verified within 90 days of the dataset date.

Missing values are not zero

Unknown, unpublished, inapplicable, or unconfirmed values are shown as “Not listed,” never as a zero price. A zero appears only when the official source explicitly says the charge is free.

• Cached input not listed: 28 records
• Batch input not listed: 46 records
• Batch output not listed: 46 records

Special and conditional pricing

For context-dependent prices, the prompt-token threshold selects a tier. Cached prompt tokens are included when the published rule says so, and a whole-request rule applies the selected tier to every input, cached-input, and output token in that request—not only the tokens beyond the threshold.

Cache-write charges remain separated by published lifetime. Batch prices remain separate from standard prices. Storage, document-page, request, session-duration, and other non-token charges keep their native billing units.

Published promotional charges and list charges remain separate. A promotion without a dated transition is not used by default calculators unless a saved workload explicitly supplies the price basis.

Which models receive pricing pages

A model receives an indexed pricing page only when it is active, safe for default calculations, directly priced with finite text-token input and output rates, fully verified, linked to an official source, dated, and not an alias or a non-text-output model.

50 public records qualify and 39 do not. Records that fail the predicate are not indexed as model pages.

• 3 — non-text output (image/audio/video)
• 29 — not safe for default calculations
• 4 — preview
• 2 — status limited
• 1 — status specialized

How scenario costs are calculated

cost = tokens ÷ 1,000,000 × price per 1,000,000, calculated separately for each applicable component.

Published scenario assumptions

• Side project: 1,000 requests/month; 800,000 input tokens and 300,000 output tokens.
• Side project: 10,000 requests/month; 8,000,000 input tokens and 3,000,000 output tokens.
• Side project: 100,000 requests/month; 80,000,000 input tokens and 30,000,000 output tokens.
• Customer support chatbot: 10,000 conversations/month; 121,650,000 input tokens and 10,800,000 output tokens.
• Agent task: 1,000 tasks/month; 12,000,000 input tokens and 3,200,000 output tokens.

Compare pages apply their displayed input:output ratios to both models and solve the point where the weighted standard input and output list prices are equal. See a worked comparison. Standard prices are used unless a section explicitly says otherwise.

Use the AI API cost calculator, retry and fallback calculator, model pricing pages, or compare pages.

What creates a price-change event

A record is created when pricing semantics or a published model lifecycle changes between trusted snapshots. Source URL edits, verification timestamp updates, metadata changes, and reordered files do not create events.

Review the model price monitor or subscribe to its Atom feed.

Corrections

Report an error through a new GitHub issue or the contact page. Include the affected page URL and an official provider source. Never send API keys or private credentials.

A corrected record receives a new verification date. When the corrected pricing semantics differ from the previous trusted record, the change also creates a price-change event.

Public data and reproducibility

The data is available from the GitHub repository, Hugging Face, the ai-api-pricing npm package, and the ai-api-pricing-mcp server. The data is licensed under CC BY 4.0; repository code is licensed under MIT.

See the dataset page, the repository methodology, and the research index.

How to cite the dataset

AICostBudget. AICostBudget AI API Pricing Dataset, version 1.0.0. https://aicostbudget.com/en/datasets/ai-api-pricing.

AICostBudget AI API pricing dataset, verified 2026-10-21, https://aicostbudget.com/en/datasets/ai-api-pricing

Limitations

AICostBudget calculations are planning estimates, not provider quotes or invoices. Provider pricing, billing rules, discounts, model availability, and account terms can change. The official provider pricing page, billing dashboard, contract, and invoice remain the final source of truth.

Regional pricing, taxes, negotiated discounts, and account-specific tiers are not modeled unless a page explicitly presents a published condition. Estimates are planning numbers, not purchase quotes.