Methodology
Token Price Index measures the user price of AI inference tokens in USD per million tokens. TPX complements it by showing token purchasing power in mTPD.
Index Design Principles
Transparent, auditable, simple, and explicit about uncertainty.
Token Price Definition
Raw provider prices are normalized to USD per 1M input and output tokens. China prices quoted in RMB are converted into USD before index calculation.
Token Price Index
TPI is the weighted average USD price for one million AI inference tokens. Higher TPI means tokens are more expensive for users.
Token Purchasing Index
TPX shows how much AI one dollar can buy. It is measured in mTPD: million tokens per USD. Higher TPX means stronger token purchasing power.
Modality Boundary
TPI and TPX are text-token indices. Video, image, audio, and multimodal generation products can affect provider-level market weight, but their task prices are not mixed into text-token price calculation unless they expose comparable input/output token pricing.
Video Models
Products such as JiMeng or SeedDance belong in a future media-generation index with units such as USD per second, USD per video, or normalized generation unit. They should not be directly averaged with USD per million text tokens.
Why Not One Company
Single-company pricing can reflect product strategy, discounts, and temporary changes. A basket reduces dependence on one provider.
Regions
The public site displays Global, US, and China indices. The data model can support additional regional baskets later.
Model Inclusion
A model can enter an index only if it has observable pricing, stable API or enterprise access, real commercial relevance, current flagship or mainline status for its provider, and a reliable source trail.
Model Exclusions
The index excludes private discounts, temporary promotions, chat-only products without API pricing, one-off benchmark releases, discontinued models, and prices that cannot be normalized into USD per 1M tokens.
Watchlist
New or newly detected models go to a watchlist first. They are reviewed for pricing clarity, API stability, availability, model quality, and provider relevance before entering a live basket.
Monthly Rebalancing
Basket membership and weights are reviewed on a monthly schedule. Source-page updates can trigger review, but they do not automatically rewrite constituents or weights.
Frontier Standard
A frontier model has stable public or enterprise API access, leading real-task performance, high-end capability in at least three of reasoning, code, long context, agents, and multimodal work, observable pricing, and stable service capacity. It is not merely a benchmark-optimized release.
Frontier Handling
Frontier classification is reviewed using capability, availability, pricing observability, and service stability. Future versions may add benchmark, reliability, and market adoption data.
Weighting Logic
Weights are updated from market-share signals first. If stable API market-share data is unavailable, the system falls back to auditable revenue or adoption proxies and labels them explicitly. Exact live constituent weights are not exposed as a free real-time feed.
Daily Weight Update
The scheduled weight job normalizes the latest reviewed signals to 100% for US, China, and Global baskets, then recalculates TPI and TPX. Weight changes are not based on token price changes.
Historical Movement
Confirmed TPI history comes from daily price and weight snapshots. Before enough real snapshots exist, the launch history uses a clearly labeled factor model so charts show expected sources of movement without pretending to be audited history.
Price Impact Factors
The core driver is compute affordability. Technology progress, better chips, serving efficiency, caching, and competition tend to push TPI down and TPX up. Compute scarcity, GPU bottlenecks, and heavy frontier demand can push TPI up and TPX down.
Market Share
Weights are provider-level first, then mapped to each provider's representative model. Private API volume is not directly observable for every provider, so the index separates exact market-share data from company-level market proxies, revenue proxies, and adoption proxies.
Regional Sleeves
The Global basket uses regional market sleeves before model-level normalization. Sleeve values can be updated when better global revenue or share data is available and summarized in public methodology releases.
Weight Constraints
Every index must total 100%. No provider is assigned weight purely because its price is high or low. Lower-confidence signals remain visible as proxies until replaced by stronger sources.
New Models
New models enter a watchlist for at least 30 days, then can join on the next monthly rebalance at up to 5% initial weight.
Source Page Changes
When a pricing page changes, the fetcher marks it needs_review. The review checks for new models, retired models, price changes, cached or batch price changes, and source reliability before any index change.
Large Price Drops
A one-time public price cut over 20% triggers at least 5 trading days of confirmation. Up to 40% is reflected during confirmation, then the rest is phased in over 10 trading days if stable and broadly available.
Source Reliability
official, cloud_marketplace, aggregator, manual, and estimated are supported. The index prioritizes official pricing URLs.
Missing Data
If a parser fails, the system keeps manual override values, records parser_failed, and shows uncertainty in the UI.
Anti-Manipulation
Limited promotions, private discounts, and non-general pricing are excluded from the main index.
Live Pricing
The fetcher parses official pricing pages where the page structure is stable. If a source cannot be parsed confidently, the system stores page metadata, keeps the last reviewed price, and marks the row for review.
Limitations
Some providers publish pricing in dynamic pages or regional consoles that are difficult to parse reliably. The site separates live parsed prices from reviewed manual prices and review-needed rows.