Relay Methodology
Last updated:
1. What this page is
API relays ("relays") are third-party services that forward model APIs from providers such as OpenAI, Anthropic and Google to their users. apiprice collects only reproducible facts from relays' public pages, with sources and verification times. We do not score or rank relays. Being listed is not an endorsement and does not mean a relay is safe.
2. Listing criteria
Both of: (1) an independent website; (2) appearing on at least two public relay rankings or lists. Listing does not require the relay's consent, and relays cannot pay to be listed or removed. We do not contact relays, log in, sign up, top up or call their APIs with a key. If a website fails to load for 30 consecutive days, we label it "website unreachable for 30 consecutive days" and move it to the bottom of the page; we do not delete it.
3. What we record, how, and how often
| Fact | How we get it | Frequency |
|---|---|---|
| Public list prices and multipliers | Fully render the public pricing page in a headless browser (no login, no sign-up, at most once per site per day), keep a timestamped screenshot, compare with our official price model by model | Automated change check daily; manual review weekly |
| Top-up ratio, bonus, minimum top-up | The relay's published top-up terms, quoted | Weekly |
| Operating entity | Website footer, terms of service, ICP licence number | Monthly |
| Domain registration date | Public RDAP lookup (rdap.org), CNNIC whois for .cn, with the lookup time | Monthly |
| Whether the website, pricing page and status page load | One plain HTTP request per day from our server, recording status code and time | Daily |
| Transparency checklist | Manual check of the homepage, footer, docs home and terms; each item recorded as yes / no / unverified | Monthly |
| Third-party tests | Quotes from public testing sites and lists, with their data date | Weekly |
Every fact shows a "last verified" time. Items not manually re-verified for 14 days are labelled "needs re-check".
4. How multipliers are calculated
Multiplier = the relay's public list price ÷ the official price we track for the same model (per 1M tokens), computed separately for input and output, first pricing tier only, excluding cached-input prices and limited-time discounts. USD list prices are compared with the official international USD price; CNY list prices with the official China-site CNY price. If a relay has several groups, routes or user tiers, we record only the one its pricing page shows by default and name it.
If dollar credit is bought in CNY, we also multiply by top-up ratio ÷ exchange rate ÷ (1 + bonus). We use 1 USD = 6.6849 CNY (open.er-api.com, refreshed daily). Only bonuses the relay publishes are used, and we show both "no bonus" and "top-tier bonus" values.
Example: a relay states "fixed 1:7 rate, up to 20% bonus" and official-equal model prices; after top-up the multiplier is 1.00 × 7 ÷ 6.6849 ÷ 1.2 ≈ 0.87 to 1.00 × 7 ÷ 6.6849 ≈ 1.05.
If a relay publishes only multipliers without per-model or base prices, we quote them and do not convert. If prices are only visible after logging in, they are shown as unverified.
A multiplier reflects list prices only. It does not tell you whether you actually get that model or whether requests succeed.
5. Why there are no scores, and how entries are ordered
What matters most about a relay is whether it serves the real model (rather than a cheaper substitute) and whether it stays available. Both require paid testing, and software-only detection is not fully reliable (see section 7). We have not done that testing. Scoring on price, domain age and other public information would push "cheap but possibly problematic" relays to the top, so we publish no overall score and no ranking.
Entries are sorted alphabetically by name. The order has nothing to do with price, quality or sponsorship. The transparency checklist shows each item only; it is never added up into a score.
6. What we have not tested
- Model authenticity (substitution, downgrading, quantization): not tested
- API uptime, error rate, latency: not tested (we only check whether the website loads; we do not call APIs)
- Refunds and support: not tested
- How data and keys are handled: not tested
These items are always shown as "not tested".
7. Security note
When you use a relay, your prompts, code, files and API keys pass through a third party. The 2026 paper "Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain" (arXiv:2604.08407, CCS 2026) tested 428 LLM API routers: 1 of 28 paid routers and 8 of 400 free routers injected malicious code into responses, and 17 used AWS canary credentials the researchers had planted (https://arxiv.org/abs/2604.08407). Another paper finds that software-only methods cannot reliably detect model substitution (arXiv:2504.04715, https://arxiv.org/abs/2504.04715).
Do not send sensitive data or production keys through a relay you do not know; review commands before letting a coding tool run them; rotate keys regularly.
8. Compliance note
Reselling official APIs may breach model providers' terms, and a relay may be cut off by its upstream at any time:
- OpenAI Services Agreement §3.1: customers may not resell or lease access to their account; §3.3(g): may not buy, sell or transfer API keys with a third party; §16.12: may not access or offer access to the services outside supported countries and territories (https://openai.com/policies/services-agreement/). Mainland China and Hong Kong are not on OpenAI's supported list (https://help.openai.com/en/articles/5347006-openai-api-supported-countries-and-territories).
- Anthropic Commercial Terms D.4: may not resell the services except as expressly approved by Anthropic (https://www.anthropic.com/legal/commercial-terms); mainland China and Hong Kong are not on the supported list (https://www.anthropic.com/supported-countries).
- Google Gemini API Additional Terms: services may only be accessed, or API clients made available to users, within an available region (https://ai.google.dev/gemini-api/terms).
These terms bind relays and end users. We record public facts only and make no finding on whether any relay complies. This is not legal advice.
9. How we cite third-party tests
We cite only sources with a public link, quote them verbatim with their data date, and never convert them into our own score or average them. If a source has a stake in the relay it tests (for example affiliate links, or it runs a relay itself), we say so. We have not re-checked third-party data.
10. Corrections and disputes
- Relays and users can both request corrections: email data@apiprice.app with the relay name, the field to change, and a link to a public page or a screenshot.
- Relays should write from an address on the same domain as their website, or publish the same statement on their website or official social account, so we can confirm who is writing.
- We handle every request within 7 days of receipt: we update what we can verify on a public page, and reply with the reason if we cannot.
- We correct factual errors; we do not remove accurate records on request. If a relay disputes a fact we cannot verify either way, it may send a statement of up to 200 words, which we attach verbatim under that item, labelled "Relay's statement".
- Every correction is logged publicly on the relay's page (time, field, before, after, reason). We do not publish the requester's name or email.
- The same process applies to errors anywhere on this page.
11. Independence
We have no commercial relationship with the listed relays and use no affiliate links or invite codes. Sponsorship does not affect any relay's listing, order or the facts we show; sponsors are listed under the same method and labelled "Sponsored" on their entry. See our independence & sponsorship policy.
12. Disclaimer
Information here comes from relays' public pages and public third-party data, recorded as of the verification time, and may have changed since. Being listed is not an endorsement and does not mean a relay is safe; a low multiplier does not mean good service. Whether to use any relay is your decision and your risk. Nothing here is legal, financial or security advice.