AWS Bedrock (Anthropic) API Reference

Status: Draft

Verified against official sources on: 2026-03-23

Confidence: Low. This file mixes confirmed Bedrock behavior with model and capability claims that still need point-by-point verification.

Official sources:

• Bedrock Anthropic Claude Messages API
• Bedrock InvokeModel
• Bedrock Anthropic request/response details
• Claude on Amazon Bedrock

Endpoints

Non-Streaming

POST https://bedrock-runtime.{REGION}.amazonaws.com/model/{MODEL_ID}/invoke

Streaming

POST https://bedrock-runtime.{REGION}.amazonaws.com/model/{MODEL_ID}/invoke-with-response-stream

Common Regions

Authentication

Bedrock uses AWS Signature V4 authentication (or Bearer token via API gateway). The Authorization header follows standard AWS signing.

Model IDs

Direct Model IDs

Cross-Region Inference Profiles

Prefix with region code for cross-region routing:

us.anthropic.claude-sonnet-4-20250514-v1:0
us.anthropic.claude-opus-4-6-20260217-v1:0
eu.anthropic.claude-sonnet-4-5-20250929-v1:0
global.anthropic.claude-opus-4-6-v1

Prefixes: us., eu., global.

Request/Response Differences from Direct Anthropic API

Request Body

The request body is the same as the Anthropic Messages API with these key differences:

Minimal Bedrock Request

{
  "anthropic_version": "bedrock-2023-05-31",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "Hello, Claude"
    }
  ]
}

All Supported Parameters

Same as Anthropic Messages API with anthropic_version added:

{
  "anthropic_version": "bedrock-2023-05-31",
  "max_tokens": 4096,
  "messages": [...],
  "system": "You are helpful",
  "temperature": 0.7,
  "top_p": 0.9,
  "top_k": 50,
  "stop_sequences": ["\n\nHuman:"],
  "tools": [...],
  "tool_choice": {"type": "auto"},
  "thinking": {"type": "enabled", "budget_tokens": 15000},
  "metadata": {"user_id": "user-123"}
}

Response Body

The response body is identical to the direct Anthropic API:

{
  "id": "msg_abc123",
  "type": "message",
  "role": "assistant",
  "content": [...],
  "model": "claude-sonnet-4-20250514",
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 10,
    "output_tokens": 50,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0
  }
}

Streaming Differences

Direct Anthropic API

Uses standard Server-Sent Events (SSE):

event: message_start
data: {"type": "message_start", "message": {...}}

event: content_block_delta
data: {"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "Hello"}}

AWS Bedrock

Uses AWS Binary Event Stream framing protocol:

1. Response is a stream of binary frames
2. Each frame has headers (:event-type, :content-type, :message-type)
3. The payload of each chunk event contains a base64-encoded JSON body
4. The JSON body contains the standard Anthropic SSE event data

The binary format requires special decoding (AWS SDK handles this automatically). When proxying, you must decode the binary frames and re-emit as standard SSE events.

Frame Structure

[prelude: 12 bytes]
  total_byte_length (4 bytes)
  headers_byte_length (4 bytes)
  prelude_crc (4 bytes)
[headers: variable]
  :event-type = "chunk"
  :content-type = "application/json"
  :message-type = "event"
[payload: variable]
  {"bytes": "<base64-encoded Anthropic event JSON>"}
[message_crc: 4 bytes]

Decoded payload example:

{"type": "content_block_delta", "index": 0, "delta": {"type": "text_delta", "text": "Hello"}}

Extended Thinking on Bedrock

Same as direct Anthropic API. Supports both legacy enabled and adaptive thinking:

{
  "anthropic_version": "bedrock-2023-05-31",
  "max_tokens": 16000,
  "thinking": {
    "type": "enabled",
    "budget_tokens": 10000
  },
  "messages": [...]
}

Or adaptive (for Opus 4.6):

{
  "thinking": {
    "type": "adaptive",
    "effort": "high"
  }
}

Note: When thinking is enabled, temperature must be 1.0. The API timeout is 60 minutes for Claude 3.7 Sonnet and Claude 4+ models. AWS SDK default timeout is 1 minute, so set read_timeout to at least 3600 seconds.

1M Context Window (Beta)

Claude Opus 4.6, Sonnet 4.6, Sonnet 4.5, and Sonnet 4 support 1M-token context on Bedrock. For Sonnet 4.5 and Sonnet 4, this requires the beta header:

{
  "anthropic_version": "bedrock-2023-05-31",
  "anthropic_beta": ["context-1m-2025-08-07"]
}

Tool Support

All Anthropic tool types are supported on Bedrock:

Converse API (Alternative)

AWS also offers the Converse API as a unified interface across all Bedrock models:

POST https://bedrock-runtime.{REGION}.amazonaws.com/model/{MODEL_ID}/converse
POST https://bedrock-runtime.{REGION}.amazonaws.com/model/{MODEL_ID}/converse-stream

The Converse API uses a different request/response format that is model-agnostic. AWS recommends it for new applications. However, the auto_ai_router uses the native Anthropic Messages API format via invoke / invoke-with-response-stream for full feature parity.

Parameter Constraints