Kingy AI Price Index

Kingy AI Price Index — methodology v1.0

Adopted October 19, 2026, for the October 2026 launch edition. It replaces the KAPI research beta at kingy.ai/kapi/ (release kapi-shadow-2026-08-24), which published dated price observations but withheld any official index.

What the index measures

The published list price, in US dollars, of:

1. API tokens for each leading AI developer's current flagship, mid-range and small text models.
2. Consumer subscriptions for each provider's standard paid plan and its top individual plan.

It measures what a buyer is charged at list price. It is not a quality ranking, a cost-per-task measure or a usage-weighted market average.

Editions and timing

• One edition per calendar month, identified YYYY-MM.
• From November 2026, prices are captured on the 1st of the month (UTC) and the edition is published within two days. The capture date is printed on every edition. A capture made later than the 3rd says so.
• The October 2026 launch edition was captured on October 19, 2026. Its comparison point is a reconstructed September 2026 baseline: the prices in effect on September 20, 2026, read from Internet Archive snapshots of the same official pages. The November 2026 edition therefore covers a 13-day interval and is labelled as a launch transition.
• A published edition is frozen. Between editions, a weekly check re-reads every source and lists verified changes under "Since this edition". Those interim changes enter the index only in the next edition.

Reconstructed history

Editions for January 2024 to August 2026 were reconstructed in October 2026 with this method, from Internet Archive snapshots of the providers' own pricing pages. Each uses the prices in effect at 00:00 UTC on the 1st of its month, read from the last snapshot taken at or before that moment and no more than 31 days earlier. Every price keeps its snapshot link, the SHA-256 of the archived bytes and a verbatim excerpt, and every tier assignment carries the provider's own words. Where the method needed interpreting for a past lineup (for example OpenAI's o-series reasoning models, which its pricing page listed apart from its flagship GPT models), the rule applied is written on the price.

• Subscription prices were added to these editions on November 1, 2026, reconstructed the same way, with no API price or index level changed. The standard tier is published from January 2024 and the top tier from June 2025, except December 2025 and July 2026, when fewer than three plan prices could be verified.
• xAI's X Premium+ counts as its standard plan through February 2025. That end date is Kingy's judgement, not something a stored page states. xAI has no standard price from March 2025 to May 2026, and no top price in any month.
• A tier with fewer than three verified prices is not published for that month. No API tier is published before April 2024, and the mid-range tier starts in June 2025.
• Moonshot AI's archived pages carry no prices, so it is absent before September 2026.
• Prices printed per 1,000 tokens or held in page data are converted to dollars per million tokens, and the published figures are shown beside them.
• The index base stays September 2026 = 100, so adding the history moved no published level.
• Reconstructed editions are labelled as such and are not indexed by search engines. Kingy did not publish an edition in those months.

Sources

Only the provider's own official pricing or model documentation counts. Resellers, aggregators and press coverage do not. For every observation Kingy keeps the URL, the capture time, the SHA-256 of the captured bytes and a verbatim excerpt that must contain the published numbers. If the excerpt check fails, the observation is withheld.

API basket

Providers (v1)

OpenAI, Anthropic, Google, xAI, DeepSeek, Moonshot AI, Mistral AI and Alibaba Cloud (Qwen). A provider qualifies if it sells self-serve, pay-as-you-go API access to a general-purpose text model at a published USD price. Providers are added or removed only with notice one edition ahead.

Tiers

Each provider contributes at most one model per tier. A provider with no distinct model for a tier is absent from that tier; one model never fills two tiers.

Eligibility:

• The model must be purchasable by any paid self-serve customer at a published price. Paid previews count. Experimental, free-only, waitlist, limited-availability and enterprise-only models do not.
• Specialised models (coding-only, audio, image, video, embedding, search, "deep research" or premium-compute "Pro" variants) are excluded.
• A provider that sells exactly two general-purpose models has its higher-priced model in Frontier and its lower-priced model in Small.
• The provider's own words that justify each tier assignment are captured and shown next to the model. Where an archived date has no positioning page, the model in the same product line as the next edition's constituent takes the tier (for example Claude Haiku → Small), and the page says so.

When a provider launches a successor, the successor replaces its predecessor in the next edition if it meets the rule. This is deliberate: the index tracks what each lab's current lineup costs, not what an aging model costs.

Price measured

• Standard (on-demand) processing. Not batch, flex, priority, fast or regional-residency rates.
• Uncached input and output, per million tokens.
• The shortest-context price band when prices vary by prompt length.
• Time-of-day pricing: the peak (higher) rate.
• Tools, search, grounding and storage are excluded.
• A promotional or time-limited price counts if every paid customer is charged it on the capture date. The scheduled future price is recorded and shown.

Calculation

• Blended price per model = 0.75 × input price + 0.25 × output price (a 3:1 input-to-output token mix, the common industry convention). Input and output prices are also published separately so anyone can apply another mix.
• Tier value = the median blended price of the tier's constituents, equally weighted. With an even number of constituents it is the mean of the middle two.
• A tier is published only if at least three constituents have verified prices.
• Month-on-month change = tier value ÷ previous edition's tier value − 1, rounded to one decimal place.
• Like-for-like changes: for each model present in both editions, any change in its own input or output price is listed. These separate real price moves from basket changes caused by launches.

Subscription basket

• Providers (v1): OpenAI, Anthropic, Google, xAI, Perplexity and Mistral AI. Microsoft is left out because its plan pages do not reliably publish the price, and its plans bundle Office apps with Copilot.
• Standard plan: each provider's core individual paid plan, above any low-cost entry plan (for example ChatGPT Plus rather than ChatGPT Go, Google AI Pro rather than Google AI Plus).
• Top plan: each provider's highest-priced individual consumer plan. A provider with only one paid individual plan appears only under Standard.
• Price: US list price, monthly billing, USD, before tax, bought on the web (not in a mobile app store).
• Business, team and enterprise plans are excluded.
• Tier value = median monthly price. Same three-constituent minimum and change rules as the API basket.

Revisions and corrections

Published values are never silently changed. An error is corrected with a dated correction note that keeps the original value in the data. A methodology change gets a new version number; earlier editions keep the method they were published under.

Limitations

• Tokens are not words. Tokenisers differ between providers and model generations. Anthropic, for example, says its 4.7-and-later models produce about 30% more tokens for the same text. A lower price per token can still mean a higher price per task.
• Reasoning models can emit many more output tokens per answer. The 3:1 blend does not capture that.
• List price only. Discounts, committed-use contracts, batch, caching and free tiers are excluded.
• Equal weights. The index does not weight by usage or market share; Kingy has no reliable usage data across providers.
• Subscriptions are not usage-equivalent. Plans differ in limits and features. The subscription index tracks the list price people pay, not value per dollar.

Licence and citation

The data is published under CC BY 4.0. Cite as: Kingy AI Price Index, <Month YYYY> edition, Kingy AI, https://kingy.ai/ai-price-index/YYYY-MM/.

Left out of v1, and why

• Alibaba Cloud (Qwen). Model Studio prices vary by deployment region, thinking mode and context band on one page. v1 leaves Qwen out until a single comparable list price can be read reliably.
• Microsoft (Microsoft 365 Premium / Pro). Microsoft's plan pages do not reliably publish the price in the page itself (it appeared in one of four fetches on October 19, 2026), and its plans bundle Office apps with Copilot. Left out of v1 subscriptions until a stable official price source exists.