FAQ

Answers to frequent AnyRouter questions: request privacy, model hallucinations, token usage, Claude max_tokens truncation, timeouts and API key management.

Check this document first for frequently asked questions

1. Model Usage and Security

Does anyrouter store the content of user API requests?

anyrouter does not store any request content you send through the API, nor does it log the responses returned by models. anyrouter acts solely as a relay channel: it securely forwards your requests to the corresponding model provider and returns the provider's response to you unchanged.

Why do official products like claude, gpt, and qwen return different results from the API?

The underlying model is the same — the official products add extra engineering on top (such as built-in system prompts)

The web version is like a fully furnished apartment: built-in search, memory, calculator, system prompts, and more
An API call is like a bare-shell apartment: it provides only the core capability, and developers must configure context and tools themselves

Why are GPT-5 series models not recommended for translation tools?

The GPT-5 series are reasoning models, designed for complex reasoning and structured generation — not for high-frequency real-time tasks.

Reasons:

Slower responses (more reasoning steps);
Higher token consumption (long system prompts and reasoning context);
Translation plugins may accidentally trigger safety policies.

For translation or chat scenarios, lightweight models such as GPT-5.4-mini or Gemini respond faster and more reliably.

Why does GPT-5 sometimes answer "I am GPT-4" when asked "who are you"?

This is a language-model hallucination: the model describes its own base model, origin, or capabilities incorrectly. When using LLMs such as GPT-4, GPT-5, or Claude, developers may see the model answer questions about its own identity confidently but inaccurately.

Notes:

This is not the platform deliberately modifying or swapping model output, and it is not a bait-and-switch — it is normal LLM behavior;
GPT-5 was not given the name "GPT-5" during training; the name was assigned officially after training was complete;
The model does not know its own name or knowledge cutoff. The OpenAI web version answers correctly because the web app injects a built-in system prompt; what we provide is the official API version, not the web version.
When you ask a model about its own identity directly through the API, the answer can be random and inaccurate, because the model has no genuine self-awareness.

Calls to some models (such as Gemini-3-Pro) keep timing out — what should I do?

Increase your timeout setting. Gemini-3-Pro is a large model whose reasoning involves lengthy thinking time; on complex tasks the response can take more than 30 seconds, so the common 30s default easily times out.

If you must use Gemini-3-Pro, extend the timeout accordingly.
If response speed matters more, switch to a lightweight model such as Gemini 2.0, which is better suited to short timeout settings.

Why did a single "hello" consume so many tokens?

Some third-party tools (such as Cline and Claude Code) automatically attach context or system prompts to each request, and this hidden content also counts toward token consumption.

Even if you only type "hello", the underlying request may contain a large amount of conversation history or preset text. This extra content comes from the tool, not from the anyrouter platform.

What is the rate limit for concurrent API requests?

The platform currently imposes no uniform limit on concurrent requests. If you run into concurrency issues, please contact support.

Why does the same prompt produce different results each time?

Large models generate text using probabilistic sampling (temperature, top-p, and so on), randomly choosing among several candidate tokens at each step.

For more stable results, lower the temperature or disable sampling;
Output differences can also be affected by context, system prompts, or network conditions.

Why do responses end early when using Claude models?

For Claude models, anyrouter currently supports two calling methods:

The OpenAI Chat-compatible endpoint
The native Anthropic Claude endpoint

When a Claude model is called through the OpenAI Chat-compatible endpoint, the system defaults to max_tokens=4096. If the request does not explicitly set a larger max_tokens, the model stops automatically once it reaches that cap. So a response that "ends before finishing" is usually not a model fault — it hit the default output-length limit.

How do I generate longer text?

With the OpenAI Chat-compatible endpoint, set a larger max_tokens manually, for example:

completion = client.chat.completions.create(
  model="claude-sonnet-4-6",
  max_tokens=6000,
  messages=[
    {
      "role": "assistant",
      "content": "总是用中文回复"
    },
    {
      "role": "user",
      "content": "What is the meaning of life?, over 6000 words"
    }
  ]
)

max_tokens must not exceed the maximum supported by the model. If output is still truncated after setting it, provide the model name and the complete request parameters so we can investigate further.

2. API Calls and Data

Which API endpoints are available?

We provide a unified gateway compatible with several mainstream model specifications:

OpenAI-standard endpoint: https://anyrouter.win/v1 (supports GPT and compatible models)
Claude auto-forwarding endpoint: https://anyrouter.win (compatible with the Anthropic SDK calling convention)

What data is recorded while I use the API?

We record only the usage data necessary to operate the service, including account information, call records, models used, token consumption, and payment information.

Privacy guarantees:

User inputs and model outputs are never stored
Data is used only for billing and service improvement — never for content analysis or sharing with third parties
anyrouter itself retains no specific request data; however, if an underlying cloud vendor (such as a model provider or hosting platform) records access logs for security or compliance reasons, that data is governed by their privacy policy.

3. Model Knowledge and Common Phenomena

What is an AI hallucination?

An AI hallucination occurs when a large language model generates information that is factually incorrect, unsupported, or entirely fabricated.

Possible causes:

Bias or gaps in the training data;
Overfitting of model parameters;
Randomness in the generation phase.

Hallucination is common to all large language models and is not a system fault.

Malformed requests;
Model unavailable or quota exceeded.

How do I manage API keys?

You can create, revoke, and rotate API keys in the console.

Security recommendations:

Never expose API keys in public environments;
Use a separate key for each project;
Rotate keys regularly to keep your account secure.