Bedrock: thinking.type.enabled not supported by newer models behind inference profile ARNs

Summary

Claude Code 2.1.116 sends thinking.type: "enabled" with budget_tokens in Bedrock API requests, but newer Claude models (accessed via updated AWS Bedrock inference profile ARNs) only accept thinking.type: "adaptive" with output_config.effort. This results in a 400 error, making Claude Code unusable with these models.

Error

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"\"thinking.type.enabled\" is not supported for this model. Use \"thinking.type.adaptive\" and \"output_config.effort\" to control thinking behavior."}}

Environment

• Claude Code version: 2.1.116 (latest as of 2026-04-21)
• Platform: macOS (Darwin, ARM64)
• Provider: AWS Bedrock (CLAUDE_CODE_USE_BEDROCK=1)
• Model: Set via application inference profile ARN in ANTHROPIC_MODEL

Steps to Reproduce

1. Configure Claude Code to use Bedrock with an inference profile ARN pointing to a newer Claude model
2. Set CLAUDE_CODE_USE_BEDROCK=1 in ~/.claude/settings.json
3. Launch Claude Code and send any message
4. Observe the 400 error

Expected Behavior

Claude Code should detect the model's required thinking format and send thinking.type: "adaptive" with output_config.effort instead of thinking.type: "enabled" with budget_tokens.

Workarounds Attempted

• claude --effort high — same error, the underlying request still uses thinking.type: "enabled"
• claude update — already on latest (2.1.116)
• No user-facing config exists to override the thinking type

Notes

The bundled source already contains awareness of thinking.type: "adaptive" (there is a deprecation warning for older models using enabled), but the Bedrock request path does not appear to use it. The fix likely involves updating the model-to-thinking-type mapping for Bedrock inference profiles that resolve to newer models.