Kingy AI
Explore
Work with Kingy
Menu
Historical AI Pricing

Public trust contract · Research beta

AI API Pricing methodology

This is the operational contract behind Kingy’s normalized API price observations, their Model Intelligence panels and the Workload Cost Calculator. It also defines why a source-checked price may be usable while the official KAPI index remains withheld.

Applies to release kapi-shadow-2026-08-24 · independent recomputation passed · baseline week 3 of 13

Current publication state: research beta. Ten source-checked prices are public and eight meet the current cohort rules. There is no official KAPI headline, rank or time series. The internal shadow computation is not part of this public release.

1. What this layer measures

Each observation records the posted first-party API price for one exact model endpoint and service tier. Standard input, cached input, cache-write, output, batch and long-context conditions stay separate. Currency and unit are explicit: this release uses US dollars per one million tokens.

KAPI is not a model benchmark. It does not claim that a cheaper token is a better result, or that list price equals the cost of a successful production workflow.

2. Qualification rules

A research-beta observation must establish the exact model identifier, public first-party API route, standard on-demand tier, billing unit, currency, price conditions, observation time and official source. The evidence snapshot must pass an operator review and checksum verification. Index eligibility adds regional availability and governed-cohort requirements; that is a stricter question than whether a published rate can power the calculator.

Important distinction: “The provider publishes this USD rate” can be true while “this endpoint qualifies for the US-scoped future index” remains false or unverified.

3. Normalization

Prices are normalized into named fields rather than collapsed into a single number. Standard input and output always remain distinct. Cache hits do not stand in for ordinary input. Prompt-cache writes remain separate where a provider charges for them. Batch values are published only when the official surface supports them unambiguously. A higher-rate long-context tier preserves its threshold and multiplier, including whether the multiplier applies to the whole request.

4. Workload calculation

The Workload Cost Calculator consumes this release directly. It selects the applicable standard or batch schedule, accounts for cache mix and context tiers, and then adds operational variables such as retries, accepted-result rate and human-review labour. Calculator usability does not grant index eligibility, and exploratory calculator settings do not alter the governed index method.

5. Null, withheld and zero

  • Null / not published means the source does not establish a usable value for that field.
  • Withheld means the system can compute or review a candidate, but Kingy is not publishing that official output.
  • Zero is a real numeric price. The system never converts null or withheld values into zero.

6. Region and eligibility

The current governed cohort requires evidence that the public route is available in the United States. OpenAI, Anthropic and xAI observations have supporting availability evidence in this release. The two DeepSeek observations have verified official USD list prices but no source establishing explicit US availability; they are calculator-usable and index-ineligible with reason code US_AVAILABILITY_UNVERIFIED.

7. Freshness and effective time

Every release records when evidence was generated. The runtime validates a seven-day research-beta freshness window; calculator payloads also carry warning and exclusion thresholds so stale or broken sources cannot silently remain current. A material provider change creates a new observation instead of overwriting the historical record.

8. Source and review chain

Every price observation references one or more source IDs. The public catalog exposes the official URL, publisher, assertions and evidence checksum but omits internal snapshot paths. Operator review and source verification must both pass. The sealed release was recomputed independently before integration.

9. Official index publication gate

The research pipeline can test a candidate index in private, but public output requires thirteen consecutive complete weekly baselines, passed independent recomputation, resolved material variances and separate publication authorization. Current progress is one complete week. Until the gate passes, the official headline value and unit remain null and the status is WITHHELD.

No public HTML, REST response, structured data or calculator payload may expose the shadow headline, creator minima or ranked results during this phase.

10. Corrections and versioning

Corrections are append-only. A correction must identify the release, affected observation, previous value, replacement, reason, evidence and publication time. Public consumers can pin the release ID; this page, the REST data, Model Intelligence panels and calculator currently share kapi-shadow-2026-08-24.

11. Limitations

The beta covers ten endpoints from four providers and is not a complete AI API market census. It does not include negotiated enterprise rates, marketplace markups, promotional tiers, inference hardware costs, quality-adjusted performance or total cost of ownership. Official sources can change without notice; users should follow the linked source and release timestamp for high-stakes procurement decisions.

12. Public interfaces