Review what's new and limitations in AI Agent Studio for the v.39 release.

What's new

Multimodal AI Skills support

AI Agent Studio now supports Multimodal AI Skills, enabling users to provide multiple input types—such as text, images, and documents—directly to AI models. This update introduces a new File input type, File prompt control, and dynamic model-based validation for supported file types and sizes. The feature is available for select models across Amazon Bedrock, Azure OpenAI, Google Vertex AI, and OpenAI. Refer the notes below for the specific model names.

Skills
New models for Amazon Bedrock Model connections

We have expanded the available AI models you can use in our product. When you create a model connection using Amazon Bedrock , you can then choose the following new models:

  • Standard: Amazon Nova Lite, Amazon Nova Micro, Amazon Nova Pro.
  • Fine tuned: Amazon Nova Lite, Amazon Nova Micro, Amazon Nova Pro.
  • Grounded by knowledge base: Amazon Nova Lite, Amazon Nova Micro, Amazon Nova Pro.
Deprecations:

The older Amazon Bedrock models (Titan Text G1 - Express, Titan Text G1 - Lite, Titan Text Premier, Anthropic Claude 3 Haiku v1, Anthropic Claude 3 Sonnet v1, Anthropic Claude Instant v1, Anthropic Claude 2.0 and Anthropic Claude 2.1 have been deprecated and disabled from selection in the Model connection interface.

Create and manage Model connections
New models for Azure OpenAI Model connections

We have expanded the available AI models you can use in our product. When you create a model connection using Azure OpenAI , you can then choose the following new models:

  • Standard: GPT-5, GPT-5 mini.
  • Grounded by AI Search: GPT-4.1.
Deprecations:

The older Azure OpenAI models (GPT-3.5 Turbo (16k) and GPT-4 (32k) have been deprecated and disabled from selection in the Model connection interface.

Create and manage Model connections
New models for Google Vertex AI Model connections

We have expanded the available AI models you can use in our product. When you create a model connection using Google Vertex AI , you can then choose the following new models:

  • Standard: Anthropic Claude 3.5 Sonnet, Anthropic Claude 3.7 Sonnet, Anthropic Claude Sonnet 4,Gemini 2.5 Pro, Gemini 2.5 Flash.
  • Fine tuned: Gemini 2.5 Flash,Gemini 2.5 Pro.
  • Grounded by data store: Gemini 2.5 Flash,Gemini 2.5 Pro.
Deprecations:

The older Google Vertex AI models (text-bison, text-bison 32k, text-unicorn , Gemini-1.0 Pro, Gemini-1.5 Flash) have been deprecated and disabled from selection in the Model connection interface.

Create and manage Model connections
New models for OpenAI Model connections

We have expanded the available AI models you can use in our product. When you create a model connection using OpenAI , you can then choose the following new models:

  • Standard: GPT-5, GPT-5 mini.
  • Fine tuned: GPT-4.1.
Deprecations:

The older OpenAI models (GPT-3.5 Turbo (16k)) have been deprecated and disabled from selection in the Model connection interface.

Create and manage Model connections

Limitations

AI Skills- Multimodal model support
No API Task support
API Tasks do not support multimodal model execution because file streams are not yet supported with the AI Skills.
No user-defined model support
User-defined models are not supported for multimodal use cases. These models handle only request and response schemas and do not process file metadata.
File size validation not implemented
File size validation during LLM prompt execution is not implemented in the current release.
Grounded connections are not supported for select Anthropic Sonnet models
Grounding is not supported for the specific Anthropic Sonnet models deployed through Amazon Bedrock because inference-based deployments do not expose a compatible model ARN. Affected models include::
  • Claude 3.5 Sonnet v2
  • Claude 3.7 Sonnet
  • Claude 4 Sonnet
  • Claude Sonnet 4.5
GPT-4o-mini Fine-Tuned model does not support Multimodal inputs
The OpenAI fine-tuned model GPT-4o-mini does not support multimodal inputs, even though the interface allow adding file variables in AI Skills.