Methodology
Every result should show its work.
Our calculator separates published generation cost from the estimated cost of obtaining a usable result.
Source priority
Official pricing pages and provider documentation come first. Unsupported or unverified configurations are excluded instead of approximated.
Raw and effective cost
Raw cost is the provider rate for one generation. Estimated effective cost multiplies that rate by the attempts-per-usable-output value you enter. That value is your workflow assumption, not our judgment of provider quality.
Ranking
Supported results are ranked by estimated monthly API cost for the submitted inputs. Affiliate relationships and sponsorships do not influence that order.
Freshness
Rates are marked for review after 14 days and stale after 30 days. Stale or invalidated rates are excluded until a human verifies the official source again. The hand-checked video rates are re-confirmed automatically every week while our tracked price data lists exactly the same price; if the data lists a different price, we show that tracked price instead.
Known limits
Current results exclude taxes, subscription bundles, purchased-credit discounts, storage, human production time, and any feature whose pricing cannot be verified from an official source.
AI API prices across providers
LLM, image and video API prices on our comparison pages are collected automatically from open, community-maintained price datasets (LiteLLM, with models.dev filling gaps; both MIT licensed) and matched across providers. Each price links to the provider's official pricing page. Resellers and API gateways are not listed; a new provider appears only after we classify it.
The datasets are checked daily. A new or changed price is published only after it appears in two consecutive daily checks; changes above 50% need seven consecutive days, zero prices and sudden tenfold jumps are treated as data errors and not published, and a model that disappears from the data stays listed for seven days before it is removed. If the daily check stops for more than 14 days, these pages are labeled as paused and removed from search indexing until it runs again.
Every confirmed change is recorded in our public price change log and its RSS feed. Each price row shows the date its value last changed in our data. We show list prices for the standard tier; batch, cached, long-context and regional pricing can differ, so confirm the provider's page before buying.
Some providers do not publish one USD price list. Scaleway and OVHcloud list prices in euros before tax; we compare them in USD at the European Central Bank reference rate named next to each price, and show the euro price alongside. Alibaba Cloud lists different prices per region and deployment scope; we use its Singapore (International) list throughout. When a dataset switches to another region or fixes a currency, that is not a price change: we correct our data without a change-log entry, and earlier log entries that turned out to be such corrections are removed from the public log.
An LLM gets its own page only when at least three providers list it, so every page compares real alternatives. Image and video models are compared on one page per category at the provider's default size or resolution; only per-image and per-second prices are included, and size or quality variants are not averaged in.
How we rank
Finder column sorting changes display order only. Developer and provider are separate filters. With a provider or context constraint, quoted prices and provider counts come only from offers that meet both constraints. Model calculators let you choose a cached-input share and required context; providers without documented qualifying rates or context limits are excluded. Cache-write fees and long-context surcharges are not modeled.
Provider tables are ordered by blended list price per 1M tokens, assuming 3 input tokens for every output token, then by input price. The calculator on each model page re-ranks providers for the token volumes you enter. No provider pays for its position; if we ever show a sponsored placement, it will be labeled "Sponsored" and kept outside this ranking.
Recorded-audio transcription
Speech-to-text compares time-billed APIs, not meeting-app subscriptions. Prices come from the open LiteLLM dataset where it matches the provider's page; endpoints the dataset misses or prices wrongly are read from the official pricing page and labeled with their check date. Rankings estimate base API usage for your hours and requests, including documented per-request minimums and known speaker-label charges.
Accuracy is the word error rate published by the open Open ASR Leaderboard (Apache-2.0) for the exact same model: the English average by default, FLEURS for German, French, Spanish, Italian and Portuguese. We only show a score when the leaderboard evaluated the identical model version; hosted open-weight models carry the model's score, not a test of that host. Accuracy never changes the price ranking unless you sort by it, and it is not a test of your audio.
For LLMs we show the GPQA Diamond score (graduate-level science questions) from Epoch AI's Benchmarking Hub (CC BY 4.0), using only Epoch's own runs, never third-party numbers it collects. A score is shown only when Epoch evaluated the identical model; dated snapshots, distilled and instruct variants count as different models. Where Epoch ran several reasoning-effort settings we show the best one and name it, since you choose that setting in your API call. The score belongs to the model, not to a provider: a provider serving a quantized version can do worse. It never changes the price ranking unless you sort by it.