What a relay service's "rate" actually means: 1.0, 0.3 and 0.05 explained

The definition

Rate = the price you pay at the relay service ÷ the vendor's list price.

Say a model is listed at $3 per million input tokens and $15 per million output tokens. A relay service advertises "rate 0.3". Calling that model through them costs you $0.90 per million input and $4.50 per million output.

The key point: a rate is a discount on the vendor's list price, not a discount on token count.

For the same one million tokens, a service at rate 0.3 and a direct official connection differ by 70% on the bill.

The ranges you actually see

Where the very low rates come from

Low-cost supply has roughly four sources. Knowing which one you are dealing with is most of the risk assessment:

1. Resold credit for Chinese models. DeepSeek, Qwen and GLM are an order of magnitude cheaper than Claude or GPT at list price to begin with. Add volume discounts and 0.3–0.5 is entirely reasonable. This is the most common legitimate source of a low rate.
2. Resold subscription quota. Monthly plans such as Claude Max or ChatGPT Pro split into API access and sold on. The cost is amortised against the subscription fee, which allows a very low headline rate. The risk is the vendor banning the account — if that happens, your remaining credit is in limbo.
3. Reverse-engineered endpoints. Scraping the web client's endpoint and wrapping it as an API. Extremely unstable, liable to break without notice, and normally a terms-of-service violation.
4. Misrepresentation. Advertising Opus and routing to Haiku or an open-source model. This is the worst category, and it is the most common explanation for an implausibly low price.

How to judge a rate below 0.1

Not every very low rate is a scam, but you have to verify it yourself. What to do:

• Ask where the supply comes from. Resold credit, resold subscription, or reverse-engineered? Legitimate operators generally answer.
• Test the model fingerprint. Use questions only the target model can answer, or compare the token-count signature of a fixed prompt. Run the same prompt against the vendor and against the relay service and compare output token counts and style.
• Read the billing detail. After a call, check whether the usage field matches what you were charged. Some services shave the billing rate on top of the headline rate.
• Look at response behaviour. Reverse-engineered and resold-subscription routes usually swing wildly in latency and start throwing errors as soon as concurrency rises.

Three things people forget when costing with a rate

1. Output price dominates. Output tokens typically cost 4–6× input. For long-form generation, output is the bulk of the bill.
2. Cache pricing differs by an order of magnitude. Some vendors charge 10% of the standard input price on cache hits. If your workload has a fixed system prompt, work out the cache benefit first — the headline rate will mislead you.
3. Tokenisers differ, so token counts differ. The same text can vary by 30% in token count between vendors. "Cheaper per unit" is not the same as "cheaper in practice".

The rate calculator on our rankings page lets you plug in your own input and output volumes

and see the gap between the vendor list price and what you would pay at a given rate.

In one line

A rate is neither true nor false on its own. The question is whether the channel behind it is legitimate.

When you see 0.05, do not think "how cheap". Think "how do they do it".