Detailed Analysis
Third-party API relay services (中转站) operating in the Chinese market face growing scrutiny over their transparency practices, particularly regarding how they communicate token-level billing details to end users. The article argues that the quality and trustworthiness of such relay services can be directly evaluated by examining their fee breakdowns — specifically whether users are shown granular data on input tokens, output tokens, cache reads, and cache writes for each interaction with the underlying AI model.
This matters because relay services serve as intermediaries between end users and providers like Anthropic, often in regions where direct API access is restricted or less convenient. These services purchase API access in bulk and resell it, meaning they have full visibility into the raw cost structure — including Anthropic's tiered pricing for prompt caching, which distinguishes between cache creation writes and cache read hits at meaningfully different price points. When a relay service obscures this breakdown and presents only aggregate charges, it creates an asymmetry of information that can facilitate overcharging or mask inefficient usage. The article's implicit argument is that a reputable relay station will surface this data openly, because doing so signals both technical competence and honest dealing.
The broader context here involves Anthropic's prompt caching feature, which allows repeated portions of a prompt (such as system instructions or large documents) to be stored temporarily and retrieved at a fraction of the cost of re-processing them. This feature is architecturally significant — it lowers costs for power users dramatically — but only if the relay service passes those savings through accurately. A relay that bundles all token usage into a single opaque number deprives users of any ability to audit whether caching discounts are being applied or whether their usage patterns are cost-efficient.
This development reflects a maturing ecosystem around frontier AI APIs in the Chinese market, where a cottage industry of relay services has emerged to handle access, currency conversion, and customer support. As Claude's capabilities have expanded and its pricing structure has grown more sophisticated — with distinct cost layers for input, output, cache creation, and cache retrieval — the technical bar for relay services to provide meaningful transparency has risen considerably. Services that cannot or do not display this granularity are increasingly being identified by technically savvy users as either incompetent or untrustworthy.
The linked Reddit image, appearing to show a detailed token usage breakdown, functions as a reference standard in this discussion — essentially a benchmark for what "doing it right" looks like. This crowdsourced form of accountability, where users share screenshots to compare relay service behavior, represents an informal but effective market mechanism for weeding out opaque intermediaries. It also indirectly reinforces the importance of Anthropic's own API documentation and pricing transparency, since the relay services that earn user trust are precisely those that faithfully reflect and communicate Anthropic's underlying cost structure rather than abstracting it away.
Read original article →