AnthropicModel could respect thought signatures generated by BedrockConverseModel with Anthropic models

[jerry-reevo]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

We're currently using AWS Bedrock as our main provider for Claude 4.5 with extended thinking enabled, however, they have not approved capacity for the 1M context window version, so we fallback on Anthropic when the context window exceeds the normal 200k tokens.

The issue, however, is when we fallback now, the Anthropic API rejects the message history with the following error (example):

pydantic_ai.exceptions.ModelHTTPError: status_code: 400, model_name: claude-sonnet-4-5-20250929, body: {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages.15.content.0.type: Expected `thinking` or `redacted_thinking`, but found `text`. When `thinking` is enabled, a final `assistant` message must start with a thinking block (preceeding the lastmost set of `tool_use` and `tool_result` blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable `thinking`. Please consult our documentation at https://docs.claude.com/en/docs/build-with-claude/extended-thinking'}, 'request_id': 'req_011CWCv4PotRxZ3b4h32C8b9'}

The preceeding error from Bedrock is:

pydantic_ai.exceptions.ModelHTTPError: status_code: 400, model_name: us.anthropic.claude-sonnet-4-5-20250929-v1:0, body: {'Error': {'Message': 'The model returned the following errors: Input is too long for requested model.', 'Code': 'ValidationException'}, 'ResponseMetadata': {'RequestId': '47d301d2-6c56-44d6-8c62-dd525f2284c6', 'HTTPStatusCode': 400, 'HTTPHeaders': {'date': 'Wed, 17 Dec 2025 20:47:35 GMT', 'content-type': 'application/json', 'content-length': '93', 'connection': 'keep-alive', 'x-amzn-requestid': '47d301d2-6c56-44d6-8c62-dd525f2284c6', 'x-amzn-errortype': 'ValidationException:http://internal.amazon.com/coral/com.amazon.bedrock/'}, 'RetryAttempts': 0}, 'message': 'The model returned the following errors: Input is too long for requested model.'}

Most confusing is the last message when I look at the run_result.all_messages() appears to be a successful completed response from Bedrock:

ModelResponse(parts=[
  ThinkingPart(content='****redacted****', signature='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', provider_name='bedrock'), 
  TextPart(content='Based on my analysis ****redacted****')
  ], 
  usage=RequestUsage(input_tokens=151045, cache_read_tokens=7159, output_tokens=951), 
  model_name='us.anthropic.claude-sonnet-4-5-20250929-v1:0', 
  timestamp=datetime.datetime(2025, 12, 17, 20, 47, 10, 400520, tzinfo=datetime.timezone.utc), 
  provider_name='bedrock', 
  provider_url='https://bedrock-runtime.us-west-2.amazonaws.com', 
  provider_details={'finish_reason': 'end_turn'}, 
  provider_response_id='5bca065d-0edc-4d1e-aa2d-08078c7bfa89', 
  finish_reason='stop', 
  run_id='aef4ca44-c0d6-4b20-a302-700b2808270e'
)

Example Code

Python, Pydantic AI & LLM client version

pydantic-ai 1.33.0
Python 3.13

[DouweM]

@jerry-reevo Thinking signatures are currently only sent back to the same provider we received them from, as e.g. Anthropic signatures would obviously not be understood by Google, but https://platform.claude.com/docs/en/build-with-claude/extended-thinking#thinking-encryption clarifies that in this case, we could actually do that:

signature values are compatible across platforms (Claude APIs, Amazon Bedrock, and Vertex AI). Values generated on one platform will be compatible with another.

That should already work if you use AnthropicModel with the AsyncAnthropicBedrock client (and then swap it out for a regular AsyncAnthropic when you exceed the context length).

We can also update BedrockConverseModel to store (and accept) thinking signatures with provider_name='anthropic' instead of only provider_name='bedrock' in this scenario, so that those signatures will also be respected by AnthropicModel. Would you be interested in contributing that?

[jerry-reevo]

@DouweM I tested the workaround using AsyncAnthropicBedrock as the client for AnthropicModel and fallback work again.

IMO this seems like a fair way to handle the situation - keep the same high level model type but different providers for max compatibility. Assuming this will be supported long term, I think this solution makes the most sense. The docs may benefit from some clarity around this.

[DouweM]

Yep we're adding this to do the docs: #3705