Errors and Debugging

For errors, OpenRouter returns a JSON response with the following shape:

1
type ErrorResponse = {
2
  error: {
3
    code: number;
4
    message: string;
5
    metadata?: Record<string, unknown>;
6
  };
7
};

The HTTP Response will have the same status code as error.code, forming a request error if:

• Your original request is invalid
• Your API key/account is out of credits

Otherwise, the returned HTTP response status will be 200 and any error occurred while the LLM is producing the output will be emitted in the response body or as an SSE data event.

Example code for printing errors in JavaScript:

1
const request = await fetch('https://openrouter.ai/...');
2
console.log(request.status); // Will be an error code unless the model started processing your request
3
const response = await request.json();
4
console.error(response.error?.status); // Will be an error code
5
console.error(response.error?.message);

Error Codes

• 400: Bad Request (invalid or missing params, CORS)
• 401: Invalid credentials (OAuth session expired, disabled/invalid API key)
• 402: Your account or API key has insufficient credits. Add more credits and retry the request.
• 403: Forbidden (insufficient permissions, guardrail block, or moderation flag)
• 408: Your request timed out
• 429: You are being rate limited
• 502: Your chosen model is down or we received an invalid response from it
• 503: There is no available model provider that meets your routing requirements

Retry-After Header

On 429 and 503 responses, OpenRouter may include a standard HTTP Retry-After response header indicating how many seconds to wait before retrying.

1
HTTP/1.1 429 Too Many Requests
2
Retry-After: 60

The OpenAI SDK, Anthropic SDK, Vercel AI SDK, and OpenRouter SDK already respect this header for backoff. If you’re using fetch directly, honor it before retrying:

1
const res = await fetch('https://openrouter.ai/api/v1/chat/completions', { ... });
2
if (res.status === 429 || res.status === 503) {
3
  const retryAfter = Number(res.headers.get('Retry-After'));
4
  if (Number.isFinite(retryAfter) && retryAfter > 0) {
5
    await new Promise((r) => setTimeout(r, retryAfter * 1000));
6
    // retry the request
7
  }
8
}

Moderation Errors

If your input was flagged, the error.metadata will contain information about the issue. The shape of the metadata is as follows:

1
type ModerationErrorMetadata = {
2
  reasons: string[]; // Why your input was flagged
3
  flagged_input: string; // The text segment that was flagged, limited to 100 characters. If the flagged input is longer than 100 characters, it will be truncated in the middle and replaced with ...
4
  provider_name: string; // The name of the provider that requested moderation
5
  model_slug: string;
6
};

Guardrail Errors

On inference endpoints (/chat/completions, /responses, /messages), a request can be blocked before it reaches a provider — for example by a content filter or prompt-injection detector configured via guardrails. When this happens, the response is a 403 with a message describing the block reason:

1
{
2
  "error": {
3
    "code": 403,
4
    "message": "Request blocked: prompt injection patterns detected",
5
    "metadata": {
6
      "patterns": ["ignore all previous instructions"]
7
    }
8
  }
9
}

When you opt in to router metadata via the X-OpenRouter-Metadata: enabled header, the 403 response also includes the full openrouter_metadata object with routing context and a pipeline array detailing the guardrail stages that ran:

1
{
2
  "error": {
3
    "code": 403,
4
    "message": "Request blocked: prompt injection patterns detected",
5
    "metadata": {
6
      "patterns": ["ignore all previous instructions"]
7
    }
8
  },
9
  "openrouter_metadata": {
10
    "requested": "openai/gpt-4o",
11
    "strategy": "direct",
12
    "region": "iad",
13
    "summary": "available=1",
14
    "attempt": 1,
15
    "is_byok": false,
16
    "endpoints": {
17
      "total": 1,
18
      "available": [
19
        { "provider": "OpenAI", "model": "openai/gpt-4o", "selected": false }
20
      ]
21
    },
22
    "pipeline": [
23
      {
24
        "type": "guardrail",
25
        "name": "regex_pi_detection",
26
        "guardrail_id": "grd_abc123",
27
        "guardrail_scope": "api-key",
28
        "summary": "Blocked: prompt injection detected (1 pattern matched)",
29
        "data": {
30
          "action": "blocked",
31
          "detected": true,
32
          "engines": ["regex"],
33
          "patterns": ["ignore all previous instructions"]
34
        }
35
      }
36
    ]
37
  }
38
}

The openrouter_metadata object follows the same shape as on successful responses — see Pipeline Stages for the full stage type and field reference.

Provider Errors

OpenRouter normalizes every upstream provider error into the stable, typed error_type vocabulary documented under Typed Error Codes. The same error_type values describe what went wrong whether the provider error arrives in a non-streaming response body or as a mid-stream SSE event. Native protocol codes (the Anthropic error.type, the Responses error.code) are best-effort and can differ between formats — error_type is the field to rely on across all of them.

For Chat Completions, a provider error that interrupts generation carries error_type inside error.metadata:

1
{
2
  "error": {
3
    "code": 429,
4
    "message": "Rate limit exceeded",
5
    "metadata": {
6
      "error_type": "rate_limit_exceeded",
7
      "provider_code": "rate_limited"
8
    }
9
  }
10
}

The same value is carried on mid-stream errors and on the Anthropic and Responses skins — see Skin-Specific Error Formats for the exact wire location in each format.

Masking and raw provider details

When a request fails with a 500, the message is replaced with a generic string and provider_code and openrouter_metadata are omitted, but error_type is still present (server).

For non-500 errors, the upstream provider’s own error code is surfaced in error.metadata.provider_code when available. Opt-in routing context (which provider was selected, fallback attempts, etc.) is carried in the openrouter_metadata object when the request sets X-OpenRouter-Metadata — it follows the same shape as on successful responses (routing-summary fields only; see Pipeline Stages).

When No Content is Generated

Occasionally, the model may not generate any content. This typically occurs when:

• The model is warming up from a cold start
• The system is scaling up to handle more requests

Warm-up times usually range from a few seconds to a few minutes, depending on the model and provider.

If you encounter persistent no-content issues, consider implementing a simple retry mechanism or trying again with a different provider or model that has more recent activity.

Additionally, be aware that in some cases, you may still be charged for the prompt processing cost by the upstream provider, even if no content is generated.

Streaming Error Formats

When using streaming mode (stream: true), errors are handled differently depending on when they occur:

Pre-Stream Errors

Errors that occur before any tokens are sent follow the standard error format above, with appropriate HTTP status codes. At this stage the HTTP response hasn’t been committed yet, so OpenRouter can:

• Return a proper HTTP error status (4xx/5xx)
• Silently retry with a different provider endpoint if fallback routing is enabled
• Apply rate-limit or auth checks before any work begins

You’ll see pre-stream errors for issues like invalid API keys, malformed requests, or when every available provider endpoint is exhausted before streaming starts.

Mid-Stream Errors

Once the first token has been written to the client, the HTTP 200 OK status and headers are already committed — they can’t be changed. If the provider fails at this point, OpenRouter cannot silently fail over to another provider because partial content has already been delivered to your application. The error must arrive in-band as an SSE event.

Common causes of mid-stream errors:

• Provider disconnect — the upstream connection drops after partial output (network issue, provider crash, load balancer timeout)
• Provider timeout — the model stops responding mid-generation and the read deadline expires
• Token limit hit during generation — the model reaches max_tokens or the context window fills up while producing output
• Output content filter — a content moderation system flags generated text after some of it was already streamed
• Provider overload — the upstream returns a rate-limit or capacity error after beginning to stream

Mid-stream errors are sent as Server-Sent Events (SSE) with a unified structure that includes both the error details and a completion choice:

1
type MidStreamError = {
2
  id: string;
3
  object: 'chat.completion.chunk';
4
  created: number;
5
  model: string;
6
  provider: string;
7
  error: {
8
    code: number;       // HTTP status code (e.g. 400, 429, 502)
9
    message: string;
10
    metadata?: {
11
      error_type: string;     // Typed error code — see table below
12
      provider_code?: string; // Original upstream error code (omitted on 500s)
13
    };
14
  };
15
  choices: [{
16
    index: 0;
17
    delta: { content: '' };
18
    finish_reason: 'error';
19
    native_finish_reason?: string;
20
  }];
21
};

Example SSE data:

data: {"id":"gen-abc123","object":"chat.completion.chunk","created":1234567890,"model":"openai/gpt-4o","provider":"OpenAI","error":{"code":429,"message":"Rate limit exceeded","metadata":{"error_type":"rate_limit_exceeded"}},"choices":[{"index":0,"delta":{"content":""},"finish_reason":"error"}]}

Key characteristics:

• The error appears at the top level alongside standard response fields
• error.metadata.error_type carries a typed code you can switch on programmatically — see Typed Error Codes for the full list
• A choices array is included with finish_reason: "error" to properly terminate the stream
• The HTTP status remains 200 OK since headers were already sent
• The stream is terminated after this event
• On 500-class errors, error.message is replaced with a generic string and provider_code is omitted to prevent leaking upstream details

Typed Error Codes

When a provider error reaches your application, OpenRouter tags it with a canonical error_type string — both on the non-streaming response body and on mid-stream SSE events. Use this value, not the HTTP status code alone, to programmatically distinguish error categories. It is stable across all three API skins even when the native protocol code is lossy.

Where error_type appears depends on the skin and path:

• Chat Completions: error.metadata.error_type — on the mid-stream error chunk (see Mid-Stream Errors) and on the non-streaming response when a provider error interrupts generation.
• Anthropic Messages: error.error_type on the SSE error event and the non-streaming error envelope.
• Responses: top-level error_type on the failed response, for both the streaming response.failed event and the non-streaming JSON body.

The HTTP status each error_type maps to is listed in the tables below.

Token and Length Limits

Authentication and Authorization

Rate Limiting and Availability

Request Validation

Content Policy

Image Errors

Generic

Skin-Specific Error Formats

OpenRouter exposes three API skins. Each translates the same internal provider error types into its own wire format, for non-streaming responses and in-stream errors alike. In every case error_type is the stable field; the wire location differs per skin.

Chat Completions (/api/v1/chat/completions)

Mid-stream errors appear as a chat.completion.chunk with a top-level error object (shape shown above). The error.metadata.error_type field carries the typed code.

For non-streaming requests where a provider error occurs, the error is embedded in the final response alongside any partial content:

1
{
2
  "choices": [{
3
    "message": { "role": "assistant", "content": "partial output..." },
4
    "finish_reason": "error",
5
    "error": {
6
      "code": 502,
7
      "message": "Provider disconnected mid-stream",
8
      "metadata": { "error_type": "provider_unavailable" }
9
    }
10
  }]
11
}

Responses API (/api/v1/responses)

The Responses API maps internal error types to the OpenAI Responses error code set. The mapping is narrower — many distinct internal types collapse to server_error — so the precise reason is preserved in a top-level error_type field on the response, outside the native error object:

Both the streaming terminal event and the non-streaming JSON body carry the canonical error_type at the top level of the response object. For example, an authentication failure collapses to the native server_error code but keeps error_type: "authentication":

1
{
2
  "id": "resp_abc123",
3
  "status": "failed",
4
  "error": { "code": "server_error", "message": "Invalid credentials" },
5
  "error_type": "authentication"
6
}

Streaming errors surface as one of three SSE event types, each wrapping the same response object:

1. response.failed — terminal event when the response could not complete:

   1
   {
   2
     "type": "response.failed",
   3
     "response": {
   4
       "id": "resp_abc123",
   5
       "status": "failed",
   6
       "error": { "code": "server_error", "message": "Internal server error" },
   7
       "error_type": "server"
   8
     }
   9
   }

2. response.error — error during response generation:

   1
   {
   2
     "type": "response.error",
   3
     "error": { "code": "rate_limit_exceeded", "message": "Rate limit exceeded" }
   4
   }

3. error — plain error event (matches upstream OpenAI behavior):

   1
   {
   2
     "type": "error",
   3
     "error": { "code": "invalid_api_key", "message": "Invalid API key provided" }
   4
   }

Error Code Transformations

Certain token/length errors are transformed into successful completions instead of failures:

This allows graceful handling of limit-based errors without treating them as failures.

Anthropic Messages (/api/v1/messages)

The Anthropic Messages skin maps internal types to Anthropic-native error type strings:

Because the native error.type is lossy (many internal types collapse to api_error), the canonical error_type is added inside the error object alongside it. This holds for both the non-streaming error envelope and mid-stream SSE error events.

Non-streaming error envelope:

1
{
2
  "type": "error",
3
  "error": {
4
    "type": "authentication_error",
5
    "message": "Invalid credentials",
6
    "error_type": "authentication"
7
  }
8
}

Mid-stream errors are emitted as an SSE error event with the same shape:

1
{
2
  "type": "error",
3
  "error": {
4
    "type": "overloaded_error",
5
    "message": "Provider is temporarily overloaded",
6
    "error_type": "provider_overloaded"
7
  }
8
}

Debugging

OpenRouter provides a debug option that allows you to inspect the exact request body that was sent to the upstream provider. This works with both the Chat Completions API (/api/v1/chat/completions) and the Responses API (/api/v1/responses). Useful for understanding how OpenRouter transforms your request parameters for different providers.

Debug Option Shape

The debug option is an object with the following shape:

1
type DebugOptions = {
2
  echo_upstream_body?: boolean; // If true, returns the transformed request body sent to the provider
3
};

Usage

To enable debug output, include the debug parameter in your request:

Chat Completions

1
fetch('https://openrouter.ai/api/v1/chat/completions', {
2
  method: 'POST',
3
  headers: {
4
    Authorization: 'Bearer <OPENROUTER_API_KEY>',
5
    'Content-Type': 'application/json',
6
  },
7
  body: JSON.stringify({
8
    model: 'anthropic/claude-haiku-4.5',
9
    stream: true, // Debug only works with streaming
10
    messages: [
11
      { role: 'system', content: 'You are a helpful assistant.' },
12
      { role: 'user', content: 'Hello!' },
13
    ],
14
    debug: {
15
      echo_upstream_body: true,
16
    },
17
  }),
18
});
19
20
const text = await response.text();
21
22
for (const line of text.split('\n')) {
23
  if (!line.startsWith('data: ')) continue;
24
25
  const data = line.slice(6);
26
  if (data === '[DONE]') break;
27
28
  const parsed = JSON.parse(data);
29
30
  if (parsed.debug?.echo_upstream_body) {
31
    console.log('\nDebug:', JSON.stringify(parsed.debug.echo_upstream_body, null, 2));
32
  }
33
34
  process.stdout.write(parsed.choices?.[0]?.delta?.content ?? '');
35
}

Responses API

1
fetch('https://openrouter.ai/api/v1/responses', {
2
  method: 'POST',
3
  headers: {
4
    Authorization: 'Bearer <OPENROUTER_API_KEY>',
5
    'Content-Type': 'application/json',
6
  },
7
  body: JSON.stringify({
8
    model: 'anthropic/claude-haiku-4.5',
9
    stream: true,
10
    input: 'Hello!',
11
    debug: {
12
      echo_upstream_body: true,
13
    },
14
  }),
15
});
16
17
const text = await response.text();
18
19
for (const line of text.split('\n')) {
20
  if (!line.startsWith('data: ')) continue;
21
22
  const data = line.slice(6);
23
  if (data === '[DONE]') break;
24
25
  const parsed = JSON.parse(data);
26
27
  if (parsed.type === 'response.debug') {
28
    console.log('\nDebug:', JSON.stringify(parsed.debug, null, 2));
29
  }
30
}

Debug Response Format

Chat Completions

When debug.echo_upstream_body is set to true, OpenRouter sends a debug chunk as the first chunk in the streaming response. This chunk has an empty choices array and includes a debug field with the transformed request body:

1
{
2
  "id": "gen-xxxxx",
3
  "provider": "Anthropic",
4
  "model": "anthropic/claude-haiku-4.5",
5
  "object": "chat.completion.chunk",
6
  "created": 1234567890,
7
  "choices": [],
8
  "debug": {
9
    "echo_upstream_body": {
10
      "system": [
11
        { "type": "text", "text": "You are a helpful assistant." }
12
      ],
13
      "messages": [
14
        { "role": "user", "content": "Hello!" }
15
      ],
16
      "model": "claude-haiku-4-5-20251001",
17
      "stream": true,
18
      "max_tokens": 64000,
19
      "temperature": 1
20
    }
21
  }
22
}

Responses API

On the Responses API, debug data arrives as a response.debug SSE event:

1
{
2
  "type": "response.debug",
3
  "debug": {
4
    "echo_upstream_body": {
5
      "model": "claude-haiku-4-5-20251001",
6
      "messages": [
7
        { "role": "user", "content": "Hello!" }
8
      ],
9
      "stream": true,
10
      "max_tokens": 64000,
11
      "temperature": 1
12
    }
13
  },
14
  "sequence_number": 0
15
}

Important Notes

Use Cases

The debug output is particularly useful for:

1. Understanding Parameter Transformations: See how OpenRouter maps your parameters to provider-specific formats (e.g., how max_tokens is set, how temperature is handled).

2. Verifying Message Formatting: Check how OpenRouter combines and formats your messages for different providers (e.g., how system messages are concatenated, how user messages are merged).

3. Checking Applied Defaults: See what default values OpenRouter applies when parameters are not specified in your request.

4. Debugging Provider Fallbacks: When using provider fallbacks, a debug chunk will be sent for each attempted provider, allowing you to see which providers were tried and what parameters were sent to each.

Privacy and Redaction

OpenRouter will make a best effort to automatically redact potentially sensitive or noisy data from debug output. Remember that the debug option is not intended for production.