Initial Checks
Description
The Bedrock provider does not support thinking.type: "adaptive", which AWS requires for Claude Sonnet 4.6 and Opus 4.6. The provider always sends {"type": "enabled", "budget_tokens": N}, which is deprecated for these models.
This has a concrete impact: we tested directly against the Bedrock Converse API and confirmed that with thinking.type: "enabled", thinking is silently disabled after the first generation in multi-turn agent runs. The model thinks on the first call but produces zero reasoningContent blocks on all subsequent calls (after tool use), regardless of whether thinking blocks are properly round-tripped in the message history.
Switching to thinking.type: "adaptive" restores thinking on subsequent generations:
| Call |
Thinking config |
Thinking blocks round-tripped |
Produces thinking |
| 1st |
enabled + budget_tokens |
N/A |
Yes |
| 2nd |
enabled + budget_tokens |
Yes |
No |
| 2nd |
enabled + budget_tokens |
No |
No |
| 2nd |
adaptive |
Yes |
Yes |
| 2nd |
adaptive |
No |
No |
This means any pydantic-ai user running a multi-turn agent on Bedrock with Sonnet 4.6 / Opus 4.6 silently gets no reasoning after the first LLM call — where tool results are interpreted and decisions are made.
The Anthropic direct API provider already supports adaptive thinking (#4216, #3894). The Bedrock provider does not — the unified thinking docs note "No adaptive support" for Bedrock.
Note: the effort sub-parameter inside adaptive (e.g. {"type": "adaptive", "effort": "high"}) is not yet accepted by the Bedrock Converse API — it returns ValidationException: thinking.adaptive.effort: Extra inputs are not permitted. Plain {"type": "adaptive"} works and defaults to high.
Workaround
Pass the config manually via bedrock_additional_model_requests_fields:
BedrockModelSettings(
bedrock_additional_model_requests_fields={
"thinking": {"type": "adaptive"}
}
)This works because _translate_thinking skips the thinking key when it's already present in the additional fields.
Python, Pydantic AI & LLM client version
- Python: 3.12
- Pydantic AI: 1.85.1
- boto3 / botocore (Bedrock)
Initial Checks
Description
The Bedrock provider does not support
thinking.type: "adaptive", which AWS requires for Claude Sonnet 4.6 and Opus 4.6. The provider always sends{"type": "enabled", "budget_tokens": N}, which is deprecated for these models.This has a concrete impact: we tested directly against the Bedrock Converse API and confirmed that with
thinking.type: "enabled", thinking is silently disabled after the first generation in multi-turn agent runs. The model thinks on the first call but produces zeroreasoningContentblocks on all subsequent calls (after tool use), regardless of whether thinking blocks are properly round-tripped in the message history.Switching to
thinking.type: "adaptive"restores thinking on subsequent generations:enabled+budget_tokensenabled+budget_tokensenabled+budget_tokensadaptiveadaptiveThis means any pydantic-ai user running a multi-turn agent on Bedrock with Sonnet 4.6 / Opus 4.6 silently gets no reasoning after the first LLM call — where tool results are interpreted and decisions are made.
The Anthropic direct API provider already supports adaptive thinking (#4216, #3894). The Bedrock provider does not — the unified thinking docs note "No adaptive support" for Bedrock.
Note: the
effortsub-parameter insideadaptive(e.g.{"type": "adaptive", "effort": "high"}) is not yet accepted by the Bedrock Converse API — it returnsValidationException: thinking.adaptive.effort: Extra inputs are not permitted. Plain{"type": "adaptive"}works and defaults tohigh.Workaround
Pass the config manually via
bedrock_additional_model_requests_fields:This works because
_translate_thinkingskips thethinkingkey when it's already present in the additional fields.Python, Pydantic AI & LLM client version