What KAPI is designed to measure
The planned KAPI standard-workload index measures the weekly posted cost of a fixed six-workload basket for generally available public text and code model APIs. Prices are expressed in United States dollars and normalized without silently substituting promotional, enterprise, marketplace or contract terms.
The planned headline uses the median cost of the cheapest three independent qualifying providers or creators. This prevents a single endpoint, promotion or corporate family from dominating the reading.
What qualifies
A candidate observation must establish all of the following:
- the exact public API endpoint and model identity;
- standard on-demand, synchronous access rather than a negotiated or promotional plan;
- currency, billing unit, region, tier and effective date;
- uncached input and output prices, plus separately labelled cache, batch, long-context or reasoning conditions where relevant;
- a preserved official source with observation time and checksum;
- operator review and independent recomputation.
Published price is not the same as index eligibility
A provider may publish a global or USD list price while the evidence for a KAPI-specific region, access route or effective period remains incomplete. Kingy's canonical model record can therefore show a sourced provider rate while KAPI still withholds that model from the index. The two statements answer different questions and should not be collapsed.
Example: “The provider publishes this USD rate” can be true while “this endpoint qualifies for the US-scoped KAPI cohort” remains unverified.
Unknown, withheld and zero
- Unknown means a required fact is missing or ambiguous.
- Withheld means a candidate exists but has not passed the publication gates.
- Zero is a real numeric price and must never be inferred from a blank field.
Unknown and withheld values cannot enter workload calculations, rankings or the headline index.
The six-workload basket
Each qualifying endpoint is evaluated through a fixed, versioned basket representing different input/output shapes rather than one arbitrary token ratio. Calculations use decimal arithmetic and preserve the terms that materially affect a bill. The public workload calculator can explore retries, cache usage, human review and accepted-result rate, but exploratory settings do not rewrite the governed index basket.
Baseline and release gate
The planned public headline requires thirteen consecutive complete weekly shadow releases, an independent recomputation path and no unresolved material variance. A missing qualifying week resets the evidence needed for an unbroken baseline. Until that gate passes, the correct result is no reading.
Freshness and change handling
Every accepted observation must retain its observation time and effective time. Material changes create a new observation; they do not overwrite history. Stale or superseded evidence cannot silently remain current.
Corrections
Corrections are append-only. Each correction should identify the affected release, previous value, replacement value, reason, evidence and publication time. A correction may change a historical value, but it must not erase the fact that the earlier value was published.
What KAPI does not measure
KAPI does not measure model quality, productivity, safety, actual customer invoices, negotiated discounts, market share or total value. A cheaper qualified workload is not automatically the better model for a task.
Current scope limitation
The sealed v1 file contains ten provisional endpoint identities across four providers. It is an archive of the initial public-presentation scope, not a complete market dataset and not the extensible price-observation layer the final product requires.
Return to the KAPI research preview or read all current limitations.