AWS Bedrock does not show or support Sonnet and Opus 4 Extended Thinking

[thatseant]

App Version

v3.19.1

API Provider

AWS Bedrock

Model Used

Claude Sonnet 4

🔁 Steps to Reproduce

Configration shown here: List does not showing :thinking variant

💥 Outcome Summary

List of models does not show :thinking variants for Anthropic Models. This is contrary to what was stated in #3844

Typing in various prompts after selecting Sonnet 4 also never trigger reasoning/extended thinking.

Model should either show :thinking variant or should have Enable Reasoning Mode (Similar to Anthropic as Provider) with Max Thinking Tokens configuration

📄 Relevant Logs or Errors (Optional)

[dosubot]

Hi @thatseant! I'm Dosu and I’m helping the Roo-Code team.

Based on the latest documentation and release notes, Roo Code claims to support the Anthropic Claude Sonnet 4 and Opus 4 models with the :thinking (Extended Thinking) variant for Bedrock, Anthropic, and Vertex providers, as of version 3.18 and 3.18.1 [source][source]. However, the official Bedrock provider documentation and supported model lists do not mention any :thinking variants or extended thinking capabilities for these models—only the standard model IDs like anthropic.claude-sonnet-4-20250514-v1:0 and anthropic.claude-opus-4-20250514-v1:0 are shown as available [source].

There is no documentation or configuration guidance for enabling a :thinking variant or reasoning mode for Sonnet 4 via Bedrock. The model dropdown in Roo Code for Bedrock only lists the standard variants, and there is no mention of a "Max Thinking Tokens" or "Enable Reasoning Mode" option for Bedrock as there is for Anthropic direct or Vertex AI [source].

No issues or discussions were found in the documentation about this gap, and there is no detailed documentation about the referenced pull request (#3844) beyond the release note stating support was added. The absence of the :thinking variant in your Bedrock configuration appears to be a documentation or implementation gap: the release notes claim support, but the actual Bedrock provider and UI do not expose these variants.

If you require the :thinking variant or extended reasoning features, you may need to use the Anthropic direct or Vertex AI providers, where these options are explicitly documented and available [source][source]. For Bedrock, the feature either has not been fully implemented, is not exposed by AWS, or is not yet surfaced in Roo Code's UI.

If you need further clarification or a workaround, you may want to check for updates in future releases or reach out for clarification on whether AWS Bedrock currently exposes these variants and if Roo Code plans to support them in the UI.

To reply, just mention my name @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[thatseant]

Hi I'm able to enable reasoning to work successfully in a example app that use Bedrock API with following settings:

// Claude 4 model ID (cross-region inference profile)
const CLAUDE_4_MODEL_ID = 'us.anthropic.claude-sonnet-4-20250514-v1:0';
    
const inferenceConfig = {
      maxTokens: 20000,
      temperature: 1,
      topP: 0.95
    };

// Reasoning configuration to enable thinking
const additionalModelRequestFields = {
  thinking: {
    type: "enabled",
    budget_tokens: 10000  // Budget for reasoning tokens
  }
};

const command = new ConverseStreamCommand({
  modelId: CLAUDE_4_MODEL_ID,
  messages: messages,
  inferenceConfig: inferenceConfig,
  additionalModelRequestFields: additionalModelRequestFields
}); 

https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-extended-thinking.html