Method

Where prices come from

Published price pages and public pricing endpoints, read directly from each provider and aggregator. Nothing is submitted by providers, and no listing can be bought.

How often they are re-read

Each source is re-read on its own schedule, from hourly to daily. Each offer carries the time it was last confirmed; an offer not re-confirmed for three times its source's interval (six hours at least) is flagged stale and never shown as the cheapest.

What blended means

One input-heavy request shape, 3 input tokens per 1 output token: blended = (3 × input + output) / 4, per million tokens. Input and output are always shown next to it so you can compute your own mix.

Flags

Off-peak rate marks a price from a published time-of-use schedule, with the peak rate beside it. A promotion shows the usual price at the seller and its end date when published. Price drop observed marks a lower price seen in our history, never presented as a promotion. Stale marks a price we could not re-confirm recently. Retired model id marks an id the lab now serves with another model. Prices in points, single prices for several backends and ended promotions are never shown as the cheapest. Below the lab's list price marks an offer clearly cheaper than the lab's own price with no declared discount: worth double-checking. An offer whose price changed at least 8 times in the last 7 days is marked unstable; one whose input and output prices are both more than 15 times below the median of the model's other providers (at least 3), and not yet verified, is flagged too (verified: read at the same value by two independent reads of ours). Both stay listed, with their signal, but are never selected as the cheapest. The home page features a promotion only when its price is among the three cheapest offers of its model and never above the lab's official list price.

Conditions

Training on prompts, region and sign-up terms come from each provider's published documents, with a link to the supporting document when available. When nothing is published, the condition stays unknown. We only show what is published; we cannot check it ourselves.

What is excluded

Batch and flex tiers are hidden by default, prompt-caching rates are kept separately, and none of them is folded into the headline price. Official lab list prices are kept back to 2023 for the history.

Comparable offers and the cheapest

The lowest, median and highest prices are computed on comparable offers: one price per provider (its cheapest route), standard tier, current prices only. Flex, batch and priority tiers, prices converted from points and single prices for several unnamed backends are not counted, so these prices are compared like for like. The cheapest is always the offer the API designates; it can be outside the comparable offers, and the page then says so.

Currency

Prices are in USD, converted at the ECB rate when the provider bills in another currency.

Price history

Each day at 00:00 UTC, the chart records the offer the API would have designated as the cheapest, next to the lab's official list price when it publishes one.

New models

A model is listed as new for 72 hours after it enters our tracking, if it is dated within the last 30 days by its source and has a page and a designated cheapest price; one variant per model. Order: models from a major lab first (OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, Qwen, Mistral, Moonshot, Z.ai), then the number of sellers already offering it, then our own consultations since it entered (site pages and external API or MCP calls, capped at 20 a day); the most recent wins a tie. No AI is involved. The home page shows five; the full list is on the models page.

Corrections

A wrong number is a bug. Send the model, the provider and the page you read it on to contact@inferindex.dev: it is fixed or explained.

Archived index

Each Monday's index is saved as a JSON file when it is published, and each later correction is saved as a new file: our service never overwrites an archived file. /index lists every file with its address and SHA-256 fingerprint. To check a file later, keep the fingerprint published on the day, download the file and compute its SHA-256 (sha256sum publication.json on Linux, shasum -a 256 publication.json on macOS, Get-FileHash publication.json in PowerShell). If both fingerprints match, the file is exactly the one published that day.

Self-host or API?

The /self-host answer compares the cheapest current API offer with renting GPUs by the hour. It is hardware rental only: engineering, monitoring, redundancy, start-up time, and storage and network billed separately are not counted. The break-even volume is the daily cost of the GPU configuration divided by the API price per million tokens (at a 3:1 input:output mix, unless you describe your requests).

Throughput comes from published figures only: a published measurement, an estimate scaled from one, or a published minimum (a floor: it can show that self-hosting wins, never that the API is cheaper), or the throughput you give.

The defaults (50% utilization, 70% of GPU memory for the weights, fp8, guaranteed tier) and the rule that picks the configuration are ours, and the answer says so. When the evidence is not enough, we decline to give a number rather than guess.

Full method, in English · Self-host or API?

Free models

We list the free API offers published by providers and gateways; InferIndex provides none itself. An offer is listed only when it is official (announced publicly by the provider itself or by the gateway that sells it) and has a public page we can link to. Each line shows when its terms were last checked; a line not checked again within 7 days leaves the list, as does an offer past the end date it publishes.

The main column is what the published terms say about your prompts: not used for training, not kept by default (the exceptions the provider lists are in the tooltip), kept for a number of days, used for training, or nothing published. The rate limit is the one the provider publishes; when none is published, the cell says so.

Free access comes with rate limits and data terms of its own, so a free offer is not compared with paid offers: it is never designated the cheapest offer of a model elsewhere on the site.

Free models