As checked 2026-10-25: An AI model marketed in 2026 cannot be placed in a 2021 rate card, and a page crawled today is not proof that the same rate was effective four years ago. This is a dated economics guide, not a permanent price guarantee or a personal hands-on benchmark.
Scope: global English reference; USD list prices where shown; local eligibility, currency, taxes, negotiated terms and model release dates require separate confirmation. Dynamic source-page prices have a check date, not a verified initial effective date.
Historical availability is its own data field
An AI model marketed in 2026 cannot be placed in a 2021 rate card, and a page crawled today is not proof that the same rate was effective four years ago. Keep distinct dates for model announcement, first public access, quoted price effective date, access verification, retirement and replacement. Where the vendor has released a dated PDF or archived changelog, the source can support that historical point. Where the evidence is only a dynamic pricing page, leave the original effective date null.
Use append-only price snapshots
Never overwrite a 2024 tariff row when a supplier adjusts prices in 2025 or 2026. Record a new observation with its own ID, region and unit, and connect both rows to a stable commercial SKU key. A historical question such as ‘what would 50 million input tokens have cost in June 2025?’ requires a price actually applicable in June 2025, not the most recent price of a similarly named successor model.
A defensible case study
Anthropic published a dated May 27, 2026 PDF listing Claude model SKUs, platform, pricing tier and separate cache actions. Such a dated primary artifact can support a stated effective-from snapshot, while still requiring a subsequent check to know whether the offer remained available for a particular customer on October 25. Other vendor pages can support current observations but should be flagged as dynamically maintained where they lack an explicit price-effective date.
Renames and retirements
Treat marketing aliases and immutable API model IDs as different identifiers. A developer might move from one model ID to another without any direct rate continuity. When a previously priced model is deprecated, preserve its historical row but mark the current availability unknown unless the official shutdown notice confirms a date. Do not use a successor’s features to explain the older service’s billing.
Publish uncertainty visibly
Archive pages should describe whether they contain audited invoice prices, official list-price snapshots, or merely reconstructed estimates. If exact tax, promotional eligibility or regional availability cannot be checked, do not fill gaps with assumptions. Users can still learn from a transparent comparison of billing mechanisms and documented historical changes without misleading precise numbers.
How a historical price entry is proved
A complete historical price record should ideally attach a dated publisher artifact, the exact quoted line, the model SKU, offered region, currency, unit and stated effective period. A screenshot without provenance, a reposted table or a current vendor calculator cannot prove what an API call would have cost in an earlier year. Preserve the URL and checksum of lawfully archived official evidence when licensing permits. If only an article announcing a new product survives, record that it establishes availability or announcement, not an undocumented token price.
Reporting changes without false continuity
When a service replaces one model with another, distinguish three effects: vendor list-price change for the same SKU, migration to a new model with a different meter, and workload efficiency changes due to output length or acceptance quality. Only the first is a clean tariff comparison. Historical spending can still be studied using contemporary receipts or invoices, but customer-negotiated rates should be labeled private-contract evidence, not a public universal price. The archive should answer what is demonstrably known, and explicitly isolate what remains disputed or unknown.
Decision checklist
- Identify the specific plan or API endpoint, model version, region and currency before pricing a job.
- Measure input, output, caching, generated media and failed-job units separately.
- Check the official pricing page again before procurement, especially if the SKU has no effective-date notice.
- Keep source evidence and a timestamped price ID; do not treat unknowns as zero.
Pricing evidence boundary
This cost discussion concerns Multi-vendor. Reproduce a quote only with its exact model or plan identifier, chargeable unit, billing region, processing mode and timestamp; a changing vendor rate card is not a permanent price guarantee. If any one of these dimensions is missing, keep the associated rate and worked scenario under editorial review instead of filling the gap with an estimate.
Sources & Further Reading
- Official list-price PDF with effective date 2026-05-27 and scoped Claude API rate rows — official first-party source, accessed 2026-10-25; official price-effective date 2026-05-27.
- Cohere command-a-03-2025 token list prices — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- gpt-5.4 text-token standard list rates and threshold above 272k input tokens — official first-party source, accessed 2026-10-25; original effective date not specified in source.
- Gemini Developer API shows tier/model/context/modality-specific rates and cache storage — official first-party source, accessed 2026-10-25; original effective date not specified in source.