Anthropic Messages API (v1/messages)

Include your credentials in the Authorization header as a Bearer token.

• API key — keys start with ak_. Send as Authorization: Bearer ak_....
• JWT — obtained from POST /api/auth/login. Send as Authorization: Bearer <jwt>.
• Anonymous free tier — omit the token and set the X-Free-Tier: 1 header. Limited to 1 free request per device.

curl -X POST https://api.abliteration.ai/v1/messages \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ak_YOUR_API_KEY" \
  -d '{
    "model": "abliterated-model",
    "max_tokens": 256,
    "messages": [
      {"role": "user", "content": "Hello, world!"}
    ]
  }'

Each message object has a role ("user" or "assistant") and content.

Content can be a plain string or an array of content blocks for multimodal inputs (text + images).

{
  "model": "abliterated-model",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "What is abliteration?"
    },
    {
      "role": "assistant",
      "content": "Abliteration is a technique that removes refusal vectors from LLMs."
    },
    {
      "role": "user",
      "content": "Explain how it works step by step."
    }
  ]
}

When stream is false (default), the full response is returned as JSON.

The response includes credit metering fields specific to abliteration.ai: remaining_credits, estimated_credits_used, and estimated_cost_usd.

{
  "id": "msg_abc123",
  "type": "message",
  "role": "assistant",
  "model": "abliterated-model",
  "content": [
    {
      "type": "text",
      "text": "Abliteration removes refusal vectors from language models..."
    }
  ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 42,
    "output_tokens": 128
  },
  "remaining_credits": 487,
  "estimated_credits_used": 1,
  "estimated_cost_usd": 0.00085
}

Set stream: true to receive Server-Sent Events. Each event has an event: line and a data: line with JSON.

Events follow the Anthropic streaming protocol: message_start, content_block_start, content_block_delta, content_block_stop, message_delta, and message_stop.

import anthropic

client = anthropic.Anthropic(
    base_url="https://api.abliteration.ai",
    api_key="ak_YOUR_API_KEY",
)

with client.messages.stream(
    model="abliterated-model",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Write a haiku about the ocean."}],
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

POST /v1/messages/count_tokens returns an estimated input token count without generating a response. Useful for budget checks before sending expensive prompts.

// POST /v1/messages/count_tokens
// Request — same schema as /v1/messages
{
  "model": "abliterated-model",
  "messages": [
    {"role": "user", "content": "Count these tokens."}
  ]
}

// Response
{
  "input_tokens": 12
}

Limits are enforced per user in a rolling 60-second window. Exceeding the limit returns 429 with a Retry-After header.

• API key callers: 120 requests per 60-second window.
• UI / JWT callers: 30 requests per 60-second window.
• Rate-limiter failures are fail-open — requests are allowed but the incident is logged.

Each call consumes credits based on total tokens (input + output). Credits are deducted after the response completes.

• Minimum charge is 1 credit per call.
• Credits per call: ceil(total_tokens / 500).
• Pricing: ~$5 per 1 M tokens. See the pricing page for current plans.
• Anonymous free-tier calls do not consume credits.
• If credits are insufficient, the endpoint returns 402 Permission Error.

Machine-readable specs and ready-made collections for faster integration.

• OpenAPI 3.0 spec — import into Swagger UI, Redocly, or any OpenAPI-compatible tool.
• Well-known OpenAPI discovery — /.well-known/openapi.json for automated tooling.
• Postman collection & OpenAPI guide — pre-built Postman collection with environment variables.
• Claude Code integration — use Claude Code as an agentic coding tool with the abliteration.ai backend.