Claude on AWS

Claude Fable 5.1, Fable 5, Opus 5, and Sonnet 5 are served through Amazon Bedrock. Use the same Draftworks API key and credit balance as the Azure models.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.draftworks.dev/v1",
    api_key=os.environ["DRAFTWORKS_API_KEY"],
)
response = client.chat.completions.create(
    model="claude-sonnet-5",
    max_completion_tokens=4096,
    messages=[{"role": "user", "content": "Explain this function: ..."}],
)
print(response.choices[0].message.content)

Models

Supported options

The gateway translates Chat Completions into Bedrock's native Anthropic request format. It supports system and developer instructions, user and assistant messages, function tools, tool results, tool_choice, parallel_tool_calls: false, stop sequences, and output-token limits. max_completion_tokens takes precedence over max_tokens; the default is 4,096 output tokens.

Set stream=True for OpenAI-compatible SSE. Tool argument deltas and provider token usage are included. As with Azure requests, Draftworks meters the stream independently of the client connection.

Image input uses an image_url block containing a base64 data URL for JPEG, PNG, GIF, or WebP. Remote image URLs are not fetched. Text blocks can include Anthropic cache_control fields for explicit prompt caching.

reasoning_effort accepts low, medium, high, or max where supported by the selected model. Native thinking options can be supplied through the SDK's extra_body. Model-specific combinations are validated by Bedrock. A model may reject temperature, top-p, or forced tool use when thinking is enabled.

The current AWS deployments reject schema-constrained output (response_format.type: "json_schema"). The gateway returns unsupported_parameter; use text or function tools. Legacy json_object, multiple completions, stored completions, and unsupported parameters also return errors instead of being silently ignored.

Responses can include reasoning_content and signed reasoning_details blocks from Anthropic. Preserve signed thinking blocks if sending them back in a tool conversation. These fields are extensions to the OpenAI chat format. Streams emit complete signed blocks in delta.reasoning_details when each thinking block ends. Append those blocks to the next assistant message’s reasoning_details to preserve them across tool turns.

Usage and receipts

usage.prompt_tokens includes uncached input, cache reads, and cache creation. The details report cached_tokens, cache_write_tokens, and cache_write_1h_tokens. Cache tokens are charged once at their respective rates. Output counts include tokens the provider spends on reasoning.

The API returns the public model ID. Billing events identify Amazon Bedrock as the provider, and receipts use the same signing key and verification process as Azure responses.

Region and data handling

Requests originate in AWS us-west-2 using Draftworks application inference profiles. Global profiles can process inference in another AWS region. The upstream-region header identifies the source region, not a data-residency guarantee. See data handling.

These models use POST /v1/chat/completions. Draftworks does not currently expose Anthropic's native /v1/messages API or a Responses adapter for Claude.

Fable data retention

Claude Fable 5 and Fable 5.1 require AWS provider data sharing. Prompts and completions for these models are shared with Anthropic and retained for up to 30 days for trust and safety review. store: false does not disable this retention. See AWS data retention.