Guide · 2026-11-14
How official and relay API pricing map onto each other
Per million tokens, input vs output, thinking tokens, long-context premium, cache hits — all at once.
In an API context, "how much does this model cost" is missing at least four qualifiers: per how many tokens, input or output, cache hit or not, and how long the context is. Drop any one of them and the price can differ several times over.
Align the unit first
Mainstream vendors all quote per 1M tokens, but granularity differs: OpenAI bills tokens exactly, Gemini counts thinking tokens as output, DeepSeek splits peak and off-peak. Normalise to "per 1M input tokens" and "per 1M output tokens" before comparing.
Cost input and output separately
Take Claude Sonnet 5.5: $2 input, $10 output — output costs 5× input. A request with 2,000 input and 500 output tokens costs 2000×2e-6 + 500×10e-6 = $0.009. The higher the output share, the more model choice matters.
Cache hits are the biggest lever
A repeated system prompt, a fixed knowledge base, long-document Q&A — all of them can hit the context cache. gpt-6-astra charges $1 on a cache hit against $10 for normal input (10×); DeepSeek v4-pro is $0.044 against $1.32 (30×). Structure requests well and the bill can drop by more than half.
Long-context premium
OpenAI cuts over at 272K, doubling the whole request beyond it; Gemini tiers it by model; DeepSeek replaces length premiums with peak/off-peak. Before processing long documents, confirm which tier you're in.
Why relays can't be compared by list price
A relay gives you two numbers: how much USD credit one CNY buys, and a multiplier per model. Real cost = official price × multiplier ÷ credit per CNY. A 0.3 multiplier sounds cheap, but if ¥1 only buys $0.2 of credit, your real cost is 1.5× the official price.
The calculator lives on the Relay page — two inputs and you have your answer. We don't publish relay rankings, because multipliers change weekly and a static list is a lie by the next day.