ChatBedrock

ChatBedrock(
    system_prompt=None,
    model=None,
    api=None,
    aws_profile=None,
    aws_region=None,
    base_url=None,
    max_tokens=MISSING,
    cache='auto',
    kwargs=None,
)

Chat with a model hosted on AWS Bedrock.

Bedrock exposes three APIs across two endpoints, and api selects which one to use and which request format to send:

• "responses" uses the OpenAI Responses API on the bedrock-mantle endpoint. This is the only way to reach the GPT-5 family, Grok, and Gemma on Bedrock.
• "messages" uses the Anthropic Messages API on the bedrock-mantle endpoint. Only Claude models are available here, but it includes some (like Claude Mythos) that no other Bedrock API serves.
• "converse" uses the Converse API on the bedrock-runtime endpoint. This is the default path for models available through Converse.

By default the API is picked from model, falling back to "converse" for models that aren’t recognised as mantle-only. Set api explicitly to override this – which is also how you reach a mantle model that Converse can also serve, since those aren’t auto-routed.

Note that the two endpoints have separate token quotas, so moving a model from one to the other changes which quota it consumes.

Prerequisites

NoteAWS credentials

Authentication uses botocore’s standard credential chain, so environment variables, ~/.aws/config, SSO, and instance roles all work. Pass aws_profile to select a named profile.

For direct Converse requests, pass api_key via kwargs or set AWS_BEARER_TOKEN_BEDROCK to authenticate with a bearer token instead of SigV4. An explicit api_key takes priority; otherwise an explicit aws_profile uses SigV4 before the environment bearer token is considered.

The direct Converse client uses raw HTTPX plus AWS libraries. The bedrock extra provides the AWS libraries and vendor SDKs for the mantle APIs (openai for api="responses" and anthropic for api="messages"). A chat uses one of these three API clients.

Parameters

Name	Type	Description	Default
system_prompt	Optional[str]	A system prompt to set the behavior of the assistant.	None
model	Optional[str]	The model to use for the chat. Defaults to "us.anthropic.claude-sonnet-4-6".	None
api	Optional[BedrockAPI]	Which Bedrock API to use. The default, None, picks the API from model.	None
aws_profile	Optional[str]	The AWS profile to use. Defaults to botocore’s default profile.	None
aws_region	Optional[str]	The AWS region to use. Defaults to the region from your AWS config.	None
base_url	Optional[str]	Override the endpoint URL. Needed to reach mantle’s other OpenAI-compatible path, /v1, which serves older open-weight models like gpt-oss and rejects the models /openai/v1 serves.	None
max_tokens	int | MISSING_TYPE	Maximum number of tokens to generate, defaulting to 4096 when api="converse" or api="messages". Passing this when api="responses" raises, since the Responses API has no constructor-level equivalent – set a cap per-request instead via chat.set_model_params(max_tokens=...).	MISSING
cache	Literal['auto', '5m', '1h', 'none']	Prompt caching for api="converse" and api="messages". The Responses API caches automatically, so this must be left at "auto" when api="responses".	'auto'
kwargs	Optional['OpenAIClientArgs | AnthropicClientArgs | ConverseClientArgs']	Additional client arguments. For api="converse", these are raw httpx.Client arguments (except auth and base_url, which ChatBedrock manages) and api_key is a Bedrock bearer token. For api="responses" and api="messages", these are arguments for the respective OpenAI or Anthropic SDK client. On the Converse path, api_key is consumed to create an Authorization: Bearer ... header, not passed to httpx.Client.	None

Returns

Name	Type	Description
	Chat	A Chat object.

Examples

from chatlas import ChatBedrock

# Frontier OpenAI models, which only exist on bedrock-mantle
chat = ChatBedrock(model="openai.gpt-5.6-sol")
chat.chat("What is 1 + 1? Just the number.")

# Claude through the Anthropic Messages API on mantle
chat = ChatBedrock(model="anthropic.claude-haiku-4-5", api="messages")