Converse API rejects document blocks and tool-result images for OpenAI GPT-5.6 models

Question

Two content-block shapes are rejected by the Converse API for the GPT-5.6 family even though the underlying models accept them through the Responses API. Both look like gaps in the Converse-to-OpenAI request translation rather than model capability limits, and I would like to know whether that is expected.

Environment
  - API: Converse and ConverseStream on bedrock-runtime
  - Models: openai.gpt-5.6-luna, openai.gpt-5.6-terra, openai.gpt-5.6-sol
  - Inference profiles: tested on both global.* and us.*

Issue 1: document content blocks are rejected outright

Any Converse request containing a `document` content block in user content fails validation with:
ValidationException: This model doesn't support documents

This occurs for every GPT-5.6 variant, on both global and us inference profiles, and regardless of document format or size.
The same models accept document input on the Responses API. So the capability exists at the model layer. The Converse translation appears not to map `document` blocks onto that part.
  
Issue 2: image content blocks are rejected inside toolResult

An `image` block is accepted in plain user content, which confirms image input works for these models. The same block placed inside a `toolResult` content array is rejected. This means a tool that returns visual output cannot pass that  output back to the model at all, even though the model accepts images elsewhere in the same conversation.

Questions:
  1. Is document support on Converse for the GPT-5.6 family a known gap, and is it Is the rejection of image and document blocks inside toolResult intentional for these models, or also a translation gap?
  2. If neither is planned, is the Responses API the intended path for document input on these models?

Riku_Kobayashi · Accepted Answer

Hello.

The specifications for ToolResultContentBlock state the following regarding the `image` in `toolResult.content`:  
It is unclear whether it will be supported in the future, but as things stand, I do not think it can be used with GPT-5.6.  
https://docs.aws.amazon.com/ja_jp/bedrock/latest/APIReference/API_runtime_ToolResultContentBlock.html  
> This field is only supported by Amazon Nova and Anthropic Claude 3 and 4 models.

re:Post Agent · Answer

The behavior you're observing appears to be related to how the Converse API translates requests for OpenAI GPT-5.6 models, rather than fundamental model limitations. Here's what's happening:

For document content blocks: The Converse API has specific restrictions on document support that vary by model. When you send a document block through Converse, the API validates whether the target model supports documents before translating the request. For GPT-5.6 models, this validation currently rejects document blocks even though the underlying models can process documents through the Responses API. This suggests the Converse-to-Responses translation layer hasn't been configured to map document blocks for these models yet.

For images in toolResult blocks: The Converse API does support images in user messages for GPT-5.6 models, as you've confirmed. However, the placement of content blocks matters significantly. The API has different validation rules depending on where content appears in the message structure. Images inside toolResult blocks may be subject to additional restrictions that don't apply to images in regular user content, even when the underlying model could technically process them.

Regarding your specific questions: These appear to be gaps in the Converse API's translation layer for GPT-5.6 models rather than intentional design decisions. The Converse API aims to provide a consistent interface across models, but new model families sometimes require additional work to map all capabilities through that unified interface.

For now, if you need document input or images in tool results with GPT-5.6 models, the Responses API is indeed your best path forward. The Responses API gives you direct access to the full model capabilities without the translation layer. You can continue using Converse for other parts of your workflow where these specific content types aren't needed.

I'd recommend monitoring the Amazon Bedrock documentation for updates on Converse API support for these content types with GPT-5.6 models, as these gaps are typically addressed as the service evolves.  
**Sources**  
[converse - botocore 1.43.62 documentation](https://docs.aws.amazon.com/botocore/latest/reference/services/bedrock-runtime/client/converse.html)  
[Inference using Converse API - Amazon Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html)  
[Introducing explicit prompt caching for OpenAI GPT-5.6 models on Amazon Bedrock | Artificial Intelligence](https://aws.amazon.com/blogs/machine-learning/introducing-explicit-prompt-caching-for-openai-gpt-5-6-models-on-amazon-bedrock/)

Converse API rejects document blocks and tool-result images for OpenAI GPT-5.6 models

Add your answer

Relevant content