// Models · API access
API relays vs official APIs
What relays are, the real risks (with the official terms quoted), and legitimate ways to get cheaper or unified access.
8 platforms with verified pricesofficial terms quotedchecked 2026-10-22
01Four ways to reach a model API
| Way | Who holds the key | Who sees your data | Terms |
|---|---|---|---|
| Official API | You | The model vendor | Fully allowed |
| Licensed aggregator / cloud platformOpenRouter, Alibaba Model Studio, Volcano Ark, SiliconFlow… | You (platform account) | Platform + vendor | Allowed under the platform’s agreements |
| Self-hosted gatewayLiteLLM, One API, New API, Portkey… | You (your own keys) | You + vendor | Fine for your own use |
| Unofficial relay / resold keys | The relay operator | Relay + vendor | Breaches OpenAI §3.3(g) and Anthropic resale terms |
02Risks of unofficial relays
- Terms: OpenAI forbids customers to “buy, sell, or transfer API keys from, to, or with a third party”（Services Agreement §3.3(g)） and to “Customer may not resell or lease access to its Account or any End User Account”（§3.1）；Anthropic forbids customers to “resell the Services except as expressly approved by Anthropic”（Commercial Terms）。
- Data: the relay terminates your request, so it can read and store every prompt, file and output – including secrets in code.
- Stability: upstream keys get revoked; prepaid balances can disappear with the site.
- Model swapping: you cannot verify that the model you pay for is the one that answers.
- Region rules: OpenAI forbids access from outside its supported countries (§16.12); see the official lists for OpenAI · Anthropic · Gemini.
AIMap does not list or link services that resell API keys or shared subscriptions.
03Legitimate aggregators & inference platforms
Platforms whose prices we verify (data from their official pricing pages or API). “Cheapest” = lowest blended text-model price we track there.
| Platform | Type | Models tracked | Cheapest model here | Blended $/1M | Price source |
|---|---|---|---|---|---|
| OpenRouter | Aggregator | 126 | Llama 3.1 8B Instruct | 0.025 | official ↗2026-10-22 |
| SiliconFlow | Inference platform | 25 | Ling-mini-2.0 | 0.131 | official ↗2026-10-10 |
| Together | Inference platform | 16 | DeepSeek V4 Flash 0731 | 0.175 | official ↗2026-10-10 |
| DeepInfra | Inference platform | 16 | DeepSeek-V4-Flash-0731 | 0.09 | official ↗2026-10-10 |
| Fireworks | Inference platform | 9 | NVIDIA Nemotron 3.5 Lightning 30B A3B | 0.087 | official ↗2026-10-10 |
| 阿里云百炼 | Inference platform | 6 | MiniMax-M3 (on 百炼) | 1.1 | official ↗2026-10-10 |
| Groq | Inference platform | 3 | GPT-OSS 20B | 0.131 | official ↗2026-10-10 |
| Mistral | Inference platform | 1 | GLM 5.3 (hosted by Mistral) | 2.15 | official ↗2026-10-10 |
04Chinese vendors’ official APIs (CNY pricing)
The cheapest official text model per vendor, per 1M tokens.
| Vendor | Model | Input | Output | Source |
|---|---|---|---|---|
| Alibaba (Qwen) | Qwen3.7-Flash | ¥0.2 | ¥0.8 | official ↗2026-10-10 |
| Zhipu (Z.ai) | GLM-4.6V-FlashX | ¥0.15 | ¥1.5 | official ↗2026-10-10 |
| ByteDance (Doubao) | Doubao-Seed-2.0-mini | ¥0.2 | ¥2 | official ↗2026-10-10 |
| Tencent (Hunyuan) | Hunyuan-A13B | ¥0.5 | ¥2 | official ↗2026-10-10 |
| StepFun (阶跃星辰) | Step-3.5-Flash | ¥0.7 | ¥2.1 | official ↗2026-10-10 |
| Baidu (ERNIE) | ERNIE 4.5 Turbo 128K | ¥0.8 | ¥3.2 | official ↗2026-10-10 |
| MiniMax | MiniMax-M3.1-Flash | ¥1 | ¥4 | official ↗2026-10-20 |
| DeepSeek | DeepSeek V4.1 Flash 1017 (deepseek-flash) | ¥2 | ¥8 | official ↗2026-10-19 |
| Moonshot (Kimi) | Kimi K2.6 | ¥6.5 | ¥27 | official ↗2026-10-10 |
05Self-hosted gateways (your own keys)
One OpenAI-compatible endpoint in front of all your providers, with routing, budgets and logs – without handing keys or data to a third party.
FAQ
What is an AI API relay (中转站)?
A third-party server that exposes an OpenAI-compatible endpoint and forwards your requests to upstream models (OpenAI, Claude, Gemini…) using keys or accounts it controls, charging you its own price. Technically it is a proxy; what matters is whose keys it uses and whether the upstream allows resale.
Are API relays reliable / legal?
Reselling keys breaches the upstream terms: OpenAI’s Services Agreement §3.3(g) forbids customers to “buy, sell, or transfer API keys from, to, or with a third party”, and Anthropic’s Commercial Terms forbid customers to “resell the Services except as expressly approved by Anthropic”. Keys can be revoked at any time, and the relay can read and log all your prompts and outputs. Whether it is illegal depends on your jurisdiction – this page does not give legal advice.
Why are relays so cheap?
Common reasons reported by the community include pooled or shared subscription accounts, promotional credits, keys bought in bulk under terms that forbid resale, or silently routing to a cheaper model. You usually cannot verify which one applies – that is the risk.
How do people in China use the OpenAI or Claude API officially?
OpenAI and Anthropic only serve the countries on their official lists (OpenAI: https://platform.openai.com/docs/supported-countries; Anthropic: https://www.anthropic.com/supported-countries), and OpenAI’s agreement (§16.12) forbids access from outside them. Within China, the compliant route is a domestic vendor’s official API (DeepSeek, Qwen, GLM, Kimi, Doubao…) – most offer OpenAI-compatible endpoints – or a licensed cloud platform.
Is OpenRouter a relay?
OpenRouter is a commercial aggregator with its own agreements with model providers, public per-endpoint pricing and its own terms. We track 333 OpenRouter-listed prices for 126 models. Check whether your country and use case are allowed by its terms.
Cheapest LLM APIs → · LLM API price table → · Free LLM APIs → · All AI gateways →