WaveSpeedAI

OpenRouter 403 'The request is prohibited due to a violation of provider Terms Of Service': Causes and Fixes

OpenRouter returns 403 'The request is prohibited due to a violation of provider Terms Of Service' on GPT, Claude or Gemini. How to tell an account block from a content flag, and what to do.

By WaveSpeedAI8 min read

If OpenRouter answers a harmless prompt with 403 and the message “The request is prohibited due to a violation of provider Terms Of Service”, the problem is almost never the prompt. In most public reports it is an access decision about your account, your region or your settings, and it applies to a whole provider family such as OpenAI, Anthropic or Google while other models keep working. Rewording the prompt or retrying does nothing.

This guide shows how to confirm which case you have in three requests, what you can change yourself, and how to keep your application working if the block stays.

Sources checked November 22, 2026: OpenRouter’s API documentation and Terms of Service, and public reports linked below.

What the error looks like

{
  "error": {
    "message": "The request is prohibited due to a violation of provider Terms Of Service.",
    "code": 403,
    "metadata": { "provider_name": null }
  }
}

The field to look at is metadata.provider_name. When it is null, no upstream provider is named as the source of the refusal. A team that documented its own incident in August 2026 read this as the request being rejected before a provider was chosen: every OpenAI model failed for their account, on every endpoint and with any prompt, and forcing a single provider changed nothing, while Kimi and Qwen models kept returning 200.

OpenRouter’s error reference lists 403 as “insufficient permissions, guardrail block, or moderation flag”. It does not explain this particular sentence, so what follows comes from the Terms of Service and from what users have reported.

The three causes behind it

1. Your account cannot reach that provider’s models

OpenRouter’s Terms of Service say that “Each Model Provider retains sole control over access to its Model”, and that some providers do not allow users located in certain countries or regions to access their models, in which case “you may not access these Restricted Models through our Service”. The terms also let OpenRouter suspend or limit credentials when it believes that is needed to protect the service.

In practice this shows up as: GPT, Claude and Gemini models return the 403, open-weight models do not. Since April 2026 users have reported that the block began right after paying with a card issued in mainland China and that it then followed the account to servers in other countries (one report on X). OpenRouter has not published the rule, so treat the trigger as reported, not confirmed. What matters for you is the pattern: the block is tied to the account, not to the IP address of one request.

2. A privacy or provider setting excludes the provider

Your own account settings can remove the only provider able to serve a model. A GitHub issue in an open-source project traced this same 403 to privacy and provider-routing settings, and the incident write-up above found its video models failing because an allowed-providers list did not include them. The related message, when you see it, is:

{"error": {"message": "No endpoints available matching your guardrail restrictions and data policy. Configure: https://openrouter.ai/settings/privacy", "code": 404}}

This is the cause you can fix yourself in a minute.

3. The input really was flagged

A genuine moderation flag looks different. According to the error reference, when input is flagged the error.metadata contains reasons, a truncated flagged_input, the provider_name and the model_slug. If you see those fields, the issue is the content of that one request and other prompts will work.

Find your case in three requests

Send the same trivial prompt to one closed model, a second closed model from another vendor, and one open-weight model. Use model ids from your own code.

for MODEL in openai/gpt-4o-mini anthropic/claude-haiku-4.5 moonshotai/kimi-k3; do
  echo "== $MODEL"
  curl -s https://openrouter.ai/api/v1/chat/completions \
    -H "Authorization: Bearer $OPENROUTER_API_KEY" \
    -H "Content-Type: application/json" \
    -d "{\"model\": \"$MODEL\", \"messages\": [{\"role\": \"user\", \"content\": \"Say hi\"}]}"
  echo
done
ResultMost likely causeNext step
Closed models 403 with provider_name: null, open model 200Account-level restriction on those providersContact OpenRouter support; plan a second route
404 “No endpoints available matching your guardrail restrictions and data policy”Your privacy or allowed-provider settingsReview openrouter.ai/settings/privacy and your provider allowlist
403 only on some prompts, with reasons and flagged_input in metadataContent moderation on that requestChange the input
Everything fails with 401 or 402Key or creditsCheck the key and the balance

What to do about it

Check settings first. Open the privacy settings and any allowed-providers or ignored-providers list. Loosen them, run the three requests again.

Ask OpenRouter. If closed models fail on a benign prompt and settings are not the cause, only OpenRouter can say why. Write to their support with your account email, the full error body, the time of the request and the model id. Note that under section 9 of the Terms of Service, credits left on an account are refunded on termination except when OpenRouter terminates it for a violation of the terms, so ask about your balance in the same message.

Do not try to hide your location. Reports say a proxy or an overseas server stopped helping once the restriction was applied to the account, and working around a provider’s regional rules can turn a partial block into a closed account.

Fix your error handling. The team in the August incident found their own code had shown users a “content moderation” message, because it matched the word “violation” in the 403 body. Their users were told to edit their text while the real problem was an account block. Check account-level conditions first:

def classify(status: int, body: dict) -> str:
    err = body.get("error", {})
    meta = err.get("metadata") or {}
    msg = err.get("message", "")
    if status in (401, 402):
        return "account"            # key or credits
    if status == 403 and "provider Terms Of Service" in msg and not meta.get("flagged_input"):
        return "account"            # access block: do not retry, switch route
    if status == 403 and meta.get("flagged_input"):
        return "content"            # this prompt was flagged
    if status in (408, 429, 502, 503):
        return "retry"
    return "error"

Keep a second route. For an “account” result the only thing that restores service right away is another endpoint that serves the same models. Because nearly every gateway speaks the OpenAI format, this is a base URL and a key in configuration, not a rewrite.

A second route: the WaveSpeedAI LLM API

WaveSpeedAI runs an OpenAI-compatible LLM API at https://llm.wavespeed.ai/v1. If your code already uses the OpenAI SDK against OpenRouter, switching means changing two values.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://llm.wavespeed.ai/v1",
    api_key=os.environ["WAVESPEED_API_KEY"],
)

response = client.chat.completions.create(
    model="anthropic/claude-haiku-4.5",
    messages=[{"role": "user", "content": "Say hi"}],
)
print(response.choices[0].message.content)

What you get:

  • The model families people lose on OpenRouter. The public /v1/models endpoint listed 140 language models when we checked on November 22, 2026, including models from Anthropic, OpenAI, Google, DeepSeek, Qwen, Moonshot, MiniMax and xAI. Model ids use the same vendor/model form.
  • Three API shapes. OpenAI Chat Completions and Responses, plus an Anthropic-compatible Messages endpoint, so tools built for either SDK can point at it.
  • Pay per token, no subscription. Prices per model are on the LLM page.
  • Payment options. Cards, PayPal, Alipay and WeChat Pay.
  • One account for media too. The same key and balance run the image generator and video generator models through the API.

Where it is not a match: the LLM catalog is smaller than OpenRouter’s, and it does not offer bring-your-own-key. If you depend on a niche open-weight model, check the model list first. WaveSpeedAI has its own terms of use as well; read them as you would for any provider.

For a wider comparison, see 10 Best OpenRouter Alternatives in 2026.

FAQ

What does “The request is prohibited due to a violation of provider Terms Of Service” mean on OpenRouter?

It is a 403 that says access to the requested model was refused under a provider’s terms. When metadata.provider_name is null and a harmless prompt triggers it on every model from one vendor, it is an access restriction on the account rather than a judgment about your prompt.

Why do GPT, Claude and Gemini fail while other models still work?

Model providers decide who may use their models, and OpenRouter’s terms say users in regions a provider does not serve may not access those models through OpenRouter. Open-weight models hosted by other providers are not covered by those restrictions, so they keep working.

Will a VPN, a proxy or a new API key fix the 403?

Reports say no: the restriction follows the account, so a different IP address or a new key under the same account gets the same answer.

Is it a content moderation error?

Usually not. A moderation flag includes reasons and flagged_input in the error metadata and affects only that request. The Terms Of Service 403 with provider_name: null affects every request to that provider family.

Do I lose my OpenRouter credits?

OpenRouter’s terms say remaining credits are refunded when an account is terminated, except when OpenRouter terminates it for a violation of the terms. A partial model restriction is not a termination, so your balance stays usable on the models you can still reach. Ask support for your specific case.

Share