What do LLM tokens really cost?

The Gradually LLM Price Index (GLLMPI) makes API prices comparable. It includes publicly priced text and code models that are available through an active API from their developer or an API provider. Each historical checkpoint uses its own time-specific model selection from the models available at that date. A daily run loads input and output prices from several source systems. For time-based tariffs, the regular price counts as the off-peak baseline. Input and output prices form a blended price, by default with 75% input and 25% output. The index is the popularity-weighted geometric mean of these blended prices. Price bands follow the currently displayed price. Up to $1 is low, over $1 to $5 is mid, and over $5 is high.

There is no single price for every model. One concrete example is OpenAI's GPT-6 Luna. In the official API price list checked on October 20, 2026, Standard rates for prompts with up to 272,000 input tokens are $0.10 per 1 million uncached input tokens and $0.50 per 1 million output tokens. Longer prompts use different rates. Caching, other processing modes, and tool fees can also change a bill. The index's weighted market price is a separate value that changes with the selected models, filters, and input/output mix.

The index is meant to show what the models people actually use cost. If every model counted equally, a provider with 30 small variants would have more influence than a provider with a few widely used models. If every provider counted equally, a small provider with a single model would weigh as much as OpenAI. So every model gets a weight from the Gradually popularity score. Because the score is logarithmic, 21 more points mean roughly ten times the weight. Models without a popularity score get the median weight of their provider, or the smallest weight in the index for a new provider. Each month's weights are recorded once and never changed afterwards.

With weights, a median jumps from one model to the next as soon as a few models carry a large share of the weight. The geometric mean uses every price and works with price ratios instead of absolute gaps, so a model that costs ten times as much does not pull the value up excessively. Many official consumer price indices calculate their lowest level the same way.

That depends on whether a model is new to the market. If a model was released within the 30 days before it enters the index, its price counts right away with its weight, so a cheap and widely used new model lowers the index noticeably. If a model that has been available for longer is added to the selection later, the index value does not change on that day. Only its later price changes count. The same applies to models that drop out of the selection. This way the index reflects changes in the market, not the maintenance of the model selection.

No. Your bill depends on the model, token mix, caching, batch processing, pricing tiers, possible discounts, and any request, search, or citation fees. The index is a market comparison, not an invoice estimate for one model.

The index uses the public base tier of API list prices. For time-based tariffs, that base tier is the off-peak rate. The index does not switch in real time by the clock, so current and historical points remain comparable on the same basis. It excludes context-based pricing tiers, regional price differences, negotiated discounts, caching, batch processing, request fees, search queries, citation and reasoning-token surcharges, self-hosted models, taxes, and measured market shares. The weighting uses the popularity score as an approximation of usage, not billed token volumes. Historical provider coverage is narrower before May 2024. Changes in the available model composition can also affect the series. The curated selection covers major providers, but not every API and hosting service in the market. The observed index series is a comparison benchmark, not a financial index.

The model selection comes from Gradually's central pricing database. The displayed input and output prices are loaded daily from several fixed source systems. Preview, API-only, and Pro variants are included, while retired, specialized, and legacy models are excluded.

The price bands use only the currently displayed price per 1M tokens. Prices up to $1 are in the low band. Prices above $1 and up to $5 are in the mid band. Prices above $5 are in the high band. The band updates automatically when you switch between input, output, and blended prices.

Country attribution follows the model developer’s headquarters. The country value is the popularity-weighted geometric mean of all models from that country, using the same weights as the overall index. Widely used models therefore count more than rarely used variants.

Every day. Two separate runs reload the structured input and output prices, check all source systems, and then recalculate the complete index. Nothing is published unless a price is present for every selected model and every validation passes.

The index checks official model-developer prices, model-router tariffs, and independent API catalogs. Community catalogs serve as cross-checks. The exact fixed source list and each model's source assignment are not published. The public dashboard shows the source count, checked input and output prices, selected models, update status, methodology, and validation rules.

The historical index begins on September 1, 2022. The series connects verified historical price checkpoints since September 2022 with daily market observations since August 2026. Since August 14, 2026, values have come from the automated daily collection. Earlier values are verified historical price checkpoints, not invented daily observations. Each checkpoint uses the model generation available at that date. Provider coverage is narrower before May 2024, so every data point also reports its model and provider count. Until August 13, 2026, the index is the median of provider medians, because no popularity scores exist for the historical checkpoints. Since August 14, 2026, the popularity weighting applies. Both parts are chain-linked on that day, so the line does not jump there.

The first historical checkpoint on September 1, 2022 equals 100 index points. In the history chart, a dashed line marks this starting value. A value of 50 means the price level is half of what it was then.