Skip to content
AuraMatrix
← All models

Token Price Index

Methodology

How each price gets here, what its status means, and how to fix it.

What is recorded

One record is one channel's published price for one model at one moment. Prices are in the channel's own currency per 1 million tokens. Input and output are always recorded. Cached input, cache writes, tiers and announced changes are recorded when the source lists them.

Models and channels

Each model has a stable ID such as anthropic/claude-opus-5.5. The names channels use for it are aliases. A channel is where the model can be bought: the vendor's own API (Official), a cloud platform (Cloud), a marketplace that routes to providers (Aggregator), or a vendor serving another vendor's open-weight model at its own price (Reseller).

Service plans

A channel can sell the same model under several plans: Standard (on demand), Batch, Flex, Priority (OpenAI and Anthropic call it Fast) and Ultrafast. Each plan is its own record with its own source. The list, rankings and reference price use Standard only; the other plans are on each model's page and in the API, with their difference against that channel's standard price.

Currencies

A price is recorded and shown in the currency its channel publishes it in. The index compares in US dollars: reference prices, “lowest input” and the percentages against the official price are calculated across channels that publish in USD only. Operators that run a separate China site (DeepSeek, Alibaba Bailian, Zhipu BigModel, MiniMax, Kimi) publish a price list in CNY there; those prices appear on the model page in their own section, in CNY, with the ¥ sign. They are never converted or ranked against USD prices, because exchange rates move every day and list prices do not: a converted number would show changes that no vendor made. Where an operator publishes both a USD and a CNY price for the same thing, both are recorded as published; neither is derived from the other.

Billing rules

A price is only half of a cost. The other half is how the channel applies it: above which prompt length a higher price starts and whether cached tokens count toward it, whether the whole request or only the excess is repriced, at what hours an off-peak price applies, and whether thinking tokens are charged as output. These are recorded as billing rules, separate from the price records. Each rule names the channel, models and plans it covers and includes excerpts copied word for word from the source snapshot. A program checks that every excerpt is in the snapshot whose hash is recorded, and an independent second reader re-reads the snapshot without seeing the recorded rule.

When a source does not say something, the rule says unknown. The cost estimate (/prices/v1/estimate and the estimate_cost tool) never replaces an unknown with an assumption: it returns each possible reading with its cost and states the open point. Where no rule is recorded for a channel, for example on thinking tokens, that is also unknown, not a default.

Where numbers come from

  • The public price page or machine-readable price list of each channel is fetched and stored as a snapshot.
  • The snapshot's SHA-256 hash and fetch time are shown next to every price, so anyone can tell which version of a page a number came from.
  • Prices are read from the snapshot by hand or from a structured price list. The method is shown in the evidence for each price.
  • Pages that only show prices after running scripts are listed under “Not yet covered” rather than guessed.

Daily re-check

A price read once can quietly change. So every day each source is fetched again and every published price is compared with it, one record at a time. For machine-readable price lists (OpenRouter, Azure, AWS, Google Cloud) a fixed reader extracts the price again; that reader is only used for a record after it has reproduced the record exactly from the snapshot the record was first read from. For web pages, the text that states the price is compared with the same text in the recorded snapshot: the lines that name the model, or every price-bearing line when the page is laid out differently.

  • Confirmed: the source still states the recorded prices. The time of that fetch is shown next to the price as “confirmed”.
  • Changed: a price list now states something else. A new record is added from that day's snapshot, with its hash; the old record stays. The new record is unverified until the independent second reader has re-read it. A move of more than fivefold in either direction is held for a person to check first, in case the source changed its unit.
  • Page changed: the text stating the price is different. The page is then read again by two independent AI readers from different makers. The first gets only the page's text and no tools, and must quote the lines it read the price from; a program checks that those lines are on the page and contain the reported price. The second reads the stored page file itself, in a sandbox. Neither sees the recorded price or the other's answer. If both agree with each other and with the record, a re-observation is recorded; if both agree on a new price, a new record is added; if they disagree, or more than the listed prices changed, nothing is recorded and the price is marked as not confirmed until a person has looked.
  • Source broken: the page no longer shows prices at all, because it moved or blocks the fetch. This is reported as a fault, not as “no change”.

The check is tested against itself every day before anything is published: for every page-sourced price, the recorded number is altered in a copy of the page and the check must then refuse to confirm it. Page text is treated as untrusted: an instruction written into a page is treated as data by both readers, and if one reader followed it, the other would contradict it.

Trust status

  • Verified: the price comes from the operator's own page, an independent second reader has re-read the snapshot without seeing the recorded value, and the values matched. The second reader is normally a different AI model (Codex) working blind from the snapshot file; people spot-check, and every mismatch goes to a person. The reviewer is named in the evidence for each price.
  • Listed: a marketplace (aggregator) listing that an independent second reader has re-read the same way and found to match the record. It is never promoted to Verified, because what a marketplace bills can depend on which provider it routes to; it is never the reference price.
  • Unverified: recorded but not yet re-read.

Tiers

Each model carries a tier (flagship, mid-range, small and fast) that follows the vendor's own positioning of it within its lineup. Tiers exist so that rankings compare like with like; they are an editorial grouping, not a measure of capability.

The reference price

Each model's reference is its latest official list price, preferring a verified one. Where an operator publishes more than one official price list in USD, the international list is the reference; a list for the China region is not, since that region has its own price list in CNY. Other channels are shown as a percentage above or below it. The reference carries its own status.

The tier indices (trial)

Three indices, one per tier (flagship, mid-range, small and fast), each a single number that moves only when official list prices move. They are a trial: the history is short and the rules below may still change; every rule change will be recorded here with its date.

  • Price of a model. Its official USD standard list price on the reference channel, blended 3 parts input to 1 part output, per 1 million tokens. The 3:1 mix is a stated assumption about how much a typical workload reads versus writes; it is not measured.
  • Index of a tier. Chained from day to day: yesterday's value times the geometric mean, across the models present on both days, of today's price divided by yesterday's. Every model has the same weight, because no trustworthy usage data per model exists; the geometric mean keeps one expensive model from dominating.
  • Base. 100 on the first day of recording.
  • Entries and exits. A model enters on the day of its first official record and leaves when it has none. Because only models present on both days enter the ratio, a new or withdrawn model never moves the index; only price changes do.
  • Announced changes. When a vendor publishes a future price with a date, the page shows where the index will stand on that date if nothing else moves. That figure is a projection, shown separately; the index itself changes only when the new price is recorded.
  • What it leaves out. Discounts, cached and batch prices, prices on clouds and marketplaces, and prices published in other currencies. Each value is recalculated from the records on every build; the daily values, constituents and their prices are in the JSON.

Version 0.1, first published 2026-11-11.

GPU-hour reference prices (trial)

For each GPU model, one number a day: the median, across providers, of each provider's cheapest on-demand list price for one GPU for one hour. On-demand only; spot, reserved and marketplace prices are different products and are shown next to it, never mixed in. A provider that sells whole machines is divided by the machine's GPU count, so its figure includes the CPUs, memory and disk of the machine. Every value is published with the number of providers behind it, because a provider joining or leaving the set can move a median; a GPU is shown once at least three providers list it. The change shown next to it is chained from day to day over the providers listed on both days, so a provider joining or leaving is never reported as a price move. Google Cloud prices a machine in parts (GPU, vCPU, memory); they are added up for the machine, without its local SSD. Where the provider states the machine (AWS, Azure, Google Cloud, CoreWeave), each row also shows the vCPUs and memory that come with one GPU, because the same GPU is sold in very different machines. Prices from mainland China (Alibaba Cloud, read from its own API every night) are recorded in CNY as published and shown in a section of their own; they are never converted and never enter the dollar reference prices. Recording began on 2026-11-11, so there is little history yet. The records and their sources are on the compute pages.

What counts as a change

An observed change is a difference between two successive records of the same model on the same channel read from the same source. Two records from different sources (a vendor's page and a cloud price list, say) can differ for reasons other than a price move, such as a promotion shown on one and not the other, so they are shown side by side but never reported as a change. Announced changes come from the vendor's own page with their effective date.

History and corrections

Records are never edited or deleted. A correction or a new price is a new record with its own time and source, so the history stays complete. To report an error, email hello@auramatrix.ai with the model, channel and a link to the source.

What the index covers

Models and channels from the United States and China. Since 2026-11-13 Mistral's models and its API are no longer covered: their earlier records stay in the dataset files, as every record does, but they are left out of the pages, the JSON, the MCP tools and the indices.

Update cadence

Every source is fetched once a day and every published price is compared with it (see Daily re-check). A vendor announcement is recorded within 48 hours. Each new record is re-read by an independent second reader within seven days; a price not confirmed against its source for 30 days is marked stale on the page rather than hidden.

Independence

The index is unofficial and not affiliated with any vendor. Vendor names are trademarks of their owners. No channel pays to be listed, and no relationship affects ordering, reference prices or status.

Data

The whole dataset can be downloaded as JSON, the same records as these pages. Field names and notes in the API, MCP and JSON stay in English.