abliteration.ai API errors

HTTP error codes, error body shapes (OpenAI and Anthropic), and how to handle policy-blocked requests.

abliteration.ai returns HTTP status codes with an error body that matches the upstream API’s shape — OpenAI shape on /v1/chat/completions and /v1/responses, Anthropic shape on /v1/messages.

OpenAI shape

{
  "error": {
    "message": "Invalid API key",
    "type": "invalid_request_error",
    "param": null,
    "code": "invalid_api_key"
  }
}

Anthropic shape

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "Invalid API key",
    "code": "invalid_api_key"
  }
}

Anthropic streaming error frame

When a stream fails mid-flight, the server emits an event: error SSE frame:

event: error
data: {"type": "error", "error": {"type": "api_error", "message": "..."}}

The stream then closes. Accumulate content defensively — a partial response is valid.

Status codes

Status	Cause	Fix
400	Invalid request body, unsupported model, invalid tool_choice, mixed tools + web_search_options	Check the reported param and correct the request shape
401	Missing or invalid API key	Send Authorization: Bearer $ABLIT_KEY
402	Insufficient credits	Top up in the console
403	Key lacks permission, or an enforcing policy matched	Check the error code before changing the request
429	Rate limited	Back off, honor Retry-After
503	Service unavailable, including policy_evaluation_unavailable	Retry with exponential backoff
Other 5xx	Upstream failure	Retry with exponential backoff

GET /v1/credits can return 503 credit_balance_unavailable. Honor Retry-After before retrying.

Policy errors

An enforcing policy match returns 403 policy_violation. Retrying the same input with the same policy configuration will produce the same result.

{
  "error": {
    "type": "policy_error",
    "code": "policy_violation",
    "message": "The request was blocked by a safety policy.",
    "request_id": "req_abc123"
  }
}

If a customer Block policy cannot be evaluated before its deadline, the API withholds the request or buffered response and returns the retryable error 503 policy_evaluation_unavailable:

{
  "error": {
    "type": "policy_error",
    "code": "policy_evaluation_unavailable",
    "message": "A policy could not be evaluated. Retry the request.",
    "request_id": "req_abc123"
  }
}

Customer Monitor policy failures do not block delivery. Platform protection evaluation failures fail open. A confirmed platform or customer Block match still blocks.

For streaming requests, request-policy errors are returned before the stream starts. Response-policy errors use the endpoint’s native terminal error event:

• Chat Completions emits an OpenAI error data frame and omits [DONE].
• Responses emits response.failed and omits response.completed.
• Messages emits event: error and omits message_stop.

Buffered response content and tool calls are not released before response policy evaluation completes.

Policy billing

• Requests blocked before model generation are not billed.
• Generated responses withheld by a confirmed response-policy match are billed.
• Responses withheld because a Block policy could not be evaluated are not billed.
• Responses delivered after a Monitor policy evaluation failure are billed normally.