Output truncated with finish_reason length on RelayRouter: fixing max_tokens

A response with finish_reason: "length" on RelayRouter means the model stopped because it reached the output token cap in your request, not because the gateway failed. To fix it, raise max_tokens (or maxOutputTokens for Gemini), shorten the requested output, or ask the model to continue in a follow-up turn. Check per-model details on relayrouter.io/models before choosing a value.

Why does RelayRouter return finish_reason length?

RelayRouter returns finish_reason: "length" when generation hits the output token limit that your request sets. RelayRouter passes your parameters to the upstream model using the protocol you call, so the cap you send determines where the reply stops. According to the official relayrouter.io docs, the gateway is “Compatible with the OpenAI, Anthropic and Gemini protocols”, and each protocol reports truncation in its own field. In the OpenAI format the signal is finish_reason: "length". In the Anthropic format it is stop_reason: "max_tokens". In the Gemini format it is finishReason: "MAX_TOKENS". A truncated reply is a completed response with partial text, not an error response. Common causes are a low max_tokens value copied from an example, long structured outputs such as JSON or code, and prompts that ask for many items in a single reply.

Which parameter controls output length in each protocol?

The output limit parameter depends on which of the three RelayRouter endpoints you call. All three use the same authentication header, Authorization: Bearer YOUR_API_KEY, with keys created at relayrouter.io/dashboard. The Anthropic Messages format requires max_tokens in every request, so omitting or underestimating it is a frequent source of truncation.

How do I fix truncated output without changing my code structure?

You fix truncation by changing only the length parameter in your existing SDK call. According to the official relayrouter.io/docs, you can “Keep your existing SDK, change base_url and the key, no other code changes”, so the fix stays inside the request body. For example, with the OpenAI SDK, increase the value in client.chat.completions.create(model="gpt-5.6-sol", messages=messages, max_tokens=...). With the Anthropic SDK, increase max_tokens on client.messages.create for models such as claude-opus-5-5. With Gemini, set maxOutputTokens for gemini-3.8-flash. If a single reply still ends early, append the partial answer to the conversation and ask the model to continue. Streaming is supported, which lets you watch output arrive and detect the stop reason in the final chunk.

What does a higher max_tokens cost on RelayRouter?

A higher max_tokens value allows more output tokens, and you pay for the output the model actually generates at that model's rate. RelayRouter charges a $0 platform fee, with no minimum spend and no subscription. Settlement rates differ by group: the GPT group settles at CNY 0.6 per $1 of standard usage and the Claude group at CNY 2.0 per $1, against a CNY 6.8 per $1 market reference. Some models use direct pricing, for example deepseek-v4-flash at CNY 1.1 per 1M input tokens and CNY 4.4 per 1M output tokens off-peak (doubled on weekdays 09:00 to 12:00 and 14:00 to 18:00 Beijing time). Failed or errored requests are generally not billed, but a truncated reply is a completed response, so plan your limit to avoid paying for output you then regenerate. The catalog lists about 108 models across 19 public groups on relayrouter.io/models.

FAQ

Is finish_reason length a RelayRouter error?
No. It means the model reached the output limit in your request. The response contains valid but incomplete text, and you resolve it by raising max_tokens or splitting the task.

Do I need a different SDK to change max_tokens on RelayRouter?
No. You keep your OpenAI, Anthropic or Gemini SDK, set the base URL to RelayRouter, use your RelayRouter key, and adjust the length parameter as described in relayrouter.io/docs.

Where can I check model options and rates before raising the limit?
Live per-model rates and the model list, including claude-fable-5-1, gpt-6-astra, DeepSeek, GLM, MiniMax and Moonshot models, are published at relayrouter.io/models.