Scope
This version covers pricing and service status: how providers, models, prices and plans are collected and verified, how official service status is relayed, and how relay availability is monitored. Model authenticity and billing accuracy checks will be added when they launch.
Provider types
- Official
- APIs and subscriptions run by the model maker itself.
- Platform
- Cloud providers, inference platforms and aggregators.
- Relay
- Third-party sites that resell model access.
We assign the type based on how the provider operates. An official provider and its model maker are recorded separately even when they are the same company.
Price sources
Official and platform prices are entered by hand from official pricing pages. Some overseas prices come from the open dataset models.dev and take effect only after manual review; prices of Chinese providers in CNY are always entered from their official pages. Relay prices are read from the site's public pricing endpoint and converted with the site's own ratios:
On New API sites, output and cache prices are further multiplied by their own ratios. Only groups visible without signing in are recorded, and time-of-day surcharges are recorded at the off-peak price. Prices edited by hand are never overwritten by automatic imports.
Verification date
The date next to each price is the last time it was confirmed to be current; click it to open the source. Automatic imports re-verify once a day and only update the date when the price has not changed.
Paid price
Relays list prices in site credit, and what one unit of credit actually costs differs from site to site.
When a site does not publish its top-up rate, the listed credit price is used.
Official price and ratio
Each model maker has one baseline provider, and the price in its default group is the official price of the model.
Both prices are converted to USD first and calculated separately for input and output. Tiered prices use the base tier, and per-request prices have no ratio. A ratio below 1 does not mean the same quality: model authenticity and billing accuracy are not checked yet.
Channel types
Relays often state where a group's traffic goes. We label each group with the channel type the site itself states in the group name and description; the label is the site's claim and is not verified:
- Official API
- Cloud
- Subscription pool
- Reverse-engineered
- Mixed
- Self-hosted
- Other model
Groups that the site says are answered by another model through format conversion (for example Codex converted to the Claude Code API) are labelled “Other model”: their prices are not prices of the named model, so they are listed on the provider page only and excluded from model comparisons, lowest prices and ratio statistics.
Usage estimates
Model pages estimate each provider's monthly cost from monthly input, output, cache hit rate and average input per request:
- Cached input
- Billed at the cache read price, or at the input price when no cache read price is published
- Uncached input
- Billed at the cache write price when written to cache
- Average input above a tier threshold
- The whole request uses that tier
- Per-request prices
- Requests = monthly input ÷ average input per request
Every provider uses the same cache hit rate, and top-up bonuses, coupons and plans are not included.
Exchange rates
Prices are stored in the provider's own currency. Conversion to the display currency uses the European Central Bank reference rates (fetched hourly via Frankfurter, updated on working days), and pages show the date of the rate. When a currency uses a fixed rate instead, pages mark it as a custom rate.
Change history
Changes to a price's currency, billing type or unit prices, as well as delisting and relisting, are all recorded. Changes are judged in the original currency, so exchange rate moves are not price changes.
Plans
Prices, billing periods and quota terms of coding plans, official subscriptions and relay passes are entered by hand from official pages. Quota terms are quoted verbatim and not converted to tokens.
We also record structured quotas: plans in the same provider series are compared by usage per unit of money, and request-based plans show the price per 1,000 requests; credits and multiples are defined differently by each provider and are not compared across providers.
Official service status
The status of official providers and platforms is relayed from the public data of their official status pages, synced every 5 minutes; the official pages prevail. For each provider only API-related components (such as Claude API, Responses and Chat Completions) are counted: the current status is the most severe status among them, together with any ongoing incidents.
The severity of an incident is the most severe status its affected components reached during the incident (degraded, partial outage, major outage); where a status page does not record component status (Google Cloud, Google AI Studio, xAI), the official impact level is used. The 90-day bars use one cell per day in UTC+8.
Official status pages may under-report or report late, so they show the problems the provider has confirmed; this is not an uptime figure and does not affect any score or ranking. Incident titles and updates are quoted verbatim.
Availability monitoring
With our own account, we send one minimal streaming request to a relay's API at a fixed interval (currently every 10 minutes in most cases). The prompt is a random "reply with only the number", and failures are not retried. Each probe is judged as follows:
- Success
- The end event arrives with some text or thinking content
- Failure
- Rate limits, timeouts (no first token within 60 seconds or no completion within 90 seconds), connection failures, server errors, empty responses, errors inside 200 responses and other client errors
- Not counted
- Problems with the probing account itself, such as an invalid key or low balance
Uptime is shown for the last 24 hours, 7 days and 30 days; with fewer than 20 probes we show only the count. Time to first token and total time only use successful probes and are shown as the median and P95. They depend on where the probe node sits on the network, so only compare them for the same node; pages name the node.
Corrections
If you find an error, use “Report an error” on the model, provider or plan page (sign-in required). We handle reports within 3 working days and you can see the result under My feedback; accepted fixes update the data and the verification date, and price changes appear in the change log. Suggesting a provider is free too: we verify before listing and do not accept paid listings.
Changelog
- v1.4Added the official service status method; availability monitoring only covers relays.
- v1.3Added the availability monitoring method.
- v1.2Added channel types as stated by sites, and the rule that groups answered by another model are excluded from comparisons; described how usage estimates are calculated; added quota comparisons for plans.
- v1.1Corrections are now submitted on the page, handled within 3 working days; relay prices can also come from the sub2api model plaza.
- v1.0First version, covering price sources, verification, paid prices, ratios and exchange rates.