Budgeting a chatbot: how many tokens a conversation really uses on RelayRouter

RelayRouter does not use a fixed token count per conversation. Each turn is billed on the input tokens you send (system prompt plus the conversation history resent with the request) and the output tokens the model returns. To budget, measure tokens per turn and multiply by the model's rate on relayrouter.io/models. For example, deepseek-v4-flash costs CNY 1.1 per 1M input tokens and CNY 4.4 per 1M output tokens off-peak.

Why does a conversation use more tokens with every turn?

Input tokens grow each turn because the chat endpoints are request based, so the earlier messages are sent again with every new request. On the OpenAI compatible endpoint (POST /v1/chat/completions) and the Anthropic compatible endpoint (POST /v1/messages), the messages you include in the request are what the model reads and what you pay for as input. A long system prompt is counted on every turn, and so is the full history up to that point. Output tokens, by contrast, are counted only for the reply generated in that turn. For budgeting, the total input of a conversation is the sum of every turn's input, not only the length of the final message. Trimming or summarizing old history and keeping the system prompt short are the main levers for controlling input cost.

What does each token cost on RelayRouter?

Cost per token depends on the model and, for some direct priced models, the time of day. deepseek-v4-flash is billed at 1.1x DeepSeek official time-of-day prices, and the rate doubles on weekdays from 09:00 to 12:00 and 14:00 to 18:00 Beijing time, according to relayrouter.io/models.

Both group rates are set against a CNY 6.8 per $1 market reference. There is a $0 platform fee, no minimum spend and no subscription.

How do you measure real token usage for your chatbot?

The direct method is to run a sample of real conversations through RelayRouter with your current code and record the input and output tokens per turn. RelayRouter is “Compatible with the OpenAI, Anthropic and Gemini protocols”, according to relayrouter.io, so you can test without rewriting your client. According to the relayrouter.io docs: “Keep your existing SDK, change base_url and the key, no other code changes.”

1. Create an API key at relayrouter.io/dashboard.
2. Set the base URL to https://relayrouter.io/v1 (OpenAI SDK) or https://relayrouter.io (Anthropic SDK), and send Authorization: Bearer YOUR_API_KEY.
3. Record tokens per turn, then multiply the averages by your expected conversation count and the per-token rate.

Failed or errored requests are generally not billed, so retries after errors do not normally add cost.

FAQ

Which models can I compare when budgeting? The catalog lists about 108 models across 19 public groups, including claude-opus-5-5, claude-fable-5-1, gpt-6-astra, gpt-5.6-sol, gemini-3.8-flash, plus DeepSeek, GLM, MiniMax and Moonshot.

Does running traffic off-peak reduce cost? For deepseek-v4-flash, yes: off-peak rates are half the weekday peak rates (09:00 to 12:00 and 14:00 to 18:00 Beijing time).

Are there fixed fees on top of token costs? No. RelayRouter lists a $0 platform fee, no minimum spend and no subscription, with payment by Stripe card. Live per-model rates are at relayrouter.io/models.