Reading RelayRouter usage logs to break down spend by model and key

To break down RelayRouter spend by model and key, create a separate API key for each project or service in the dashboard, record the model name each request uses, and price each model's usage with the live per-model rates at relayrouter.io/models. Group totals by key for per-team spend and by model for per-workload spend. Leave out failed or errored requests, which are generally not billed.

How do API keys help attribute spend?

Each API key is an attribution boundary, so issuing one key per project, environment or team lets you total spend per key. You create keys at relayrouter.io/dashboard, and every request authenticates with the header Authorization: Bearer YOUR_API_KEY. Assigning keys to separate services, such as a production chatbot and a staging test suite, means each key's usage reflects one workload. Switching an existing integration to a dedicated key is straightforward. According to the relayrouter.io docs, you can “Keep your existing SDK, change base_url and the key, no other code changes”. For example, a team can point its OpenAI SDK at https://relayrouter.io/v1 with one key and its Anthropic SDK at https://relayrouter.io with another, and then compare totals per key. See relayrouter.io/docs for setup details.

How is each model priced for a per-model breakdown?

Each model is priced by its group's settlement rate or its direct token price, as listed at relayrouter.io/models. The catalog has about 108 models across 19 public groups. Grouped models such as GPT and Claude are settled against standard usage in US dollars. DeepSeek's deepseek-v4-flash is billed at 1.1x DeepSeek's official time-of-day prices. Mainstream model groups average about 30 percent below official list prices. Check the live page before reconciling, because the rates shown there are the current ones.

Does the API protocol affect how spend is attributed?

No, the protocol does not change attribution, because every protocol identifies the model in the request. According to relayrouter.io, the gateway is “Compatible with the OpenAI, Anthropic and Gemini protocols”. OpenAI-compatible calls use POST /v1/chat/completions and Anthropic-compatible calls use POST /v1/messages. Both of these carry the model name in the request body. Gemini-compatible calls use POST /v1beta/models/{model}:generateContent, which carries the model name in the path. A request to gemini-3.8-flash therefore counts toward that model whichever SDK sent it. The same applies to gpt-6-astra, gpt-5.6-sol, claude-opus-5-5 and claude-fable-5-1. When you aggregate, group by model name and then map each model to its group rate. Streaming is supported on the gateway, so streamed and non-streamed requests to the same model can share one row in your breakdown.

Which requests should be excluded or adjusted?

Exclude failed or errored requests, and apply time-of-day pricing to DeepSeek usage. According to relayrouter.io, failed or errored requests are generally not billed. Your billed total should therefore reflect successful requests. There is a $0 platform fee, no minimum spend and no subscription, so no fixed charge needs to be spread across keys or models. For deepseek-v4-flash, sort each request by timestamp in Beijing time. Requests on weekdays from 09:00 to 12:00 and from 14:00 to 18:00 are billed at double the off-peak rate. Requests outside those windows use the off-peak rates of CNY 1.1 per 1M input tokens and CNY 4.4 per 1M output tokens. Payments are made by Stripe card.

FAQ

These answers cover common questions about reconciling RelayRouter spend.

Where are the current per-model rates? Live per-model rates are listed at relayrouter.io/models. They include the GPT group at CNY 0.6 and the Claude group at CNY 2.0 per $1 of standard usage.

Do I need to change code to use a separate key per project? No. You keep your existing SDK and change the base URL and the key, as described in the docs.

Are there fixed fees to allocate across keys? No. RelayRouter has a $0 platform fee, no minimum spend and no subscription, so all spend comes from usage.