[Explainer] How to Read Relay Prices: Multipliers, Official vs. Relay Costs, and Identifying “Diluted” Models | How to Read Relay Station Prices: Multipliers, Official Price Comparison, and Identifying “Diluted” Models

Technology and Creation AI
1

English

I run a small personal project (AtmoRouter, an OpenAI-compatible gateway), and the question I get most is “is this relay price actually cheap?” Here is the checklist I use myself, for any relay, not just mine.

1. What 倍率 (multiplier) really means
Most One-API / New-API style relays bill as: cost = tokens × model ratio × completion ratio × group 倍率 × (price per unit). A “0.5x” group only means 0.5 of that site’s base price, which may not equal the official price. Always convert to USD per 1M tokens, input and output separately.

2. Compare against the official price, per 1M tokens
Write down official input/output price, then the relay’s effective price. Output tokens usually dominate cost for coding agents, so a site that is cheap on input but normal on output saves less than it looks. Also check cache-read pricing if you use long contexts.

3. Check the recharge rate
“¥1 = $1 credit” vs “¥7.2 = $1” changes everything. Effective discount = (relay USD price × your CNY per credit-dollar) ÷ (official USD price × real exchange rate).

4. Spotting diluted / swapped models

  • Ask questions with known answers from the model’s knowledge cutoff, and check the reported model in the response model field.
  • Test long context: paste a ~50k token document with a hidden fact at the start and ask for it.
  • Compare tool-calling / JSON mode behaviour with the official API; weaker substitutes often break here.
  • Watch for injected text (ads, “powered by” footers) in replies; that is a red flag for anything you put in production.
  • Check latency and first-token time across a few runs; wildly inconsistent numbers can mean mixed upstreams.

5. Read the model list honestly
An OpenAI-compatible relay should expose GET /v1/models. On ours you can see it without a key:

curl -s https://atmorouter.dev/v1/models | head -c 500

Prices listed per model let you do the math above before topping up anything.

Happy to hear what checks others use. If you want to try the same checklist on AtmoRouter (180+ models), it is at atmorouter.dev, but the method works for any relay.

中文版

我自己在做一个小项目（AtmoRouter，OpenAI 兼容网关），被问最多的是“这个中转价格到底便不便宜”。下面是我自己判断任何中转站都会用的清单。

1. 倍率到底是什么
One-API / New-API 类站点大多按：费用 = tokens × 模型倍率 × 补全倍率 × 分组倍率 × 单价。“0.5 倍分组”只是该站基础价的一半，不一定等于官方价的一半。务必换算成每百万 tokens 的美元价，输入、输出分开算。

2. 对比官方价（每 1M tokens）
先记下官方输入/输出价，再算中转实际价。编程 Agent 的成本主要在输出，输入便宜但输出不便宜的站，实际省得没那么多。长上下文用户还要看缓存读取价格。

3. 看充值比例
“¥1 = $1 额度”和“¥7.2 = $1”差别巨大。实际折扣 =（中转美元价 × 每美元额度的人民币）÷（官方美元价 × 真实汇率）。

4. 识别掺水 / 换模型

  • 问知识截止附近有确定答案的问题，并看返回里的 model 字段。
  • 测长上下文：约 5 万 tokens 文档开头藏一个事实，最后让它找出来。
  • 对比工具调用 / JSON 模式表现，替代的弱模型经常在这里露馅。
  • 注意回复里有没有被注入的广告或 “powered by” 尾巴，生产环境要警惕。
  • 多跑几次看延迟和首字时间，波动特别大可能是混合上游。

5. 认真看模型列表
OpenAI 兼容的中转应提供 GET /v1/models。我们的无需 key 就能看：

curl -s https://atmorouter.dev/v1/models | head -c 500

先按上面方法算清楚再充值。

欢迎大家分享自己的检测方法。想用这套清单测一下 AtmoRouter（180+ 模型）的话在 atmorouter.dev，不过这方法对任何中转都适用。

1 Like
2

One more thing: the simplest way to spot watered-down output is to prepare a few fixed test questions of your own, run them whenever you switch providers, and compare the results. We’re the official Octomind account, offering access to multiple models. Feel free to give it a try at octomind.run