[bug] LiteLLMCallable default model is gpt-3.5-turbo, which OpenAI shuts down 2026-10-23

Describe the bug

guardrails/llm_providers.py:149 binds a model id that OpenAI has scheduled for shutdown:

145 class LiteLLMCallable(PromptCallableBase):
146     def _invoke_llm(
147         self,
148         text: Optional[str] = None,
149         model: str = "gpt-3.5-turbo",
150         messages: Optional[List[Dict]] = None,
151         *args,
152         **kwargs,
153     ) -> LLMResponse:

That bound default flows unmodified into a real provider request 59 lines later:

208         response = completion(
209             model=model,
210             *args,
211             **kwargs,
212         )

where completion is litellm.completion (imported at llm_providers.py:170).

OpenAI lists gpt-3.5-turbo for shutdown on 2026-10-23 — 36 days from today (2026-09-17). Source: https://developers.openai.com/api/docs/deprecations

After that date, any caller that lands on this default gets a hard provider error, surfaced as PromptCallableException (guardrails/classes/llm/prompt_callable.py:36).

To Reproduce

No RAIL spec needed — this is a runtime-arguments case.

from guardrails import Guard
from litellm import completion

guard = Guard()
guard(completion, messages=[{"role": "user", "content": "hello"}])   # no model=

The default is reached, not dormant:

1. guard.py:623-624 — if llm_api is not None or kwargs.get("model") is not None: api = get_llm_ask(llm_api, *args, **kwargs). The first disjunct is satisfied by llm_api=completion alone.
2. llm_providers.py:526-527 — if llm_api == completion or (llm_api is None and kwargs.get("model")): return LiteLLMCallable(*args, **kwargs). Again the first disjunct requires no model, so LiteLLMCallable is built with no "model" key in init_kwargs (get_llm_ask injects only temperature=0, line 521).
3. classes/llm/prompt_callable.py:30-34 — self._invoke_llm(*self.init_args, *args, **self.init_kwargs, **kwargs). Neither dict carries model, so Python binds the signature default.
4. Line 149 binds "gpt-3.5-turbo"; lines 208-212 send it.

I also checked that nothing upstream supplies a model first: no code under guardrails/ assigns into kwargs["model"], calls kwargs.update({"model": ...}), or uses setdefault("model", ...) on this path, and guardrails/run/ never sets a model key at all. The only other model references inside _invoke_llm are the two tracing dicts at :184 and :198, which read model but never set it. And I replicated the signature plus the dispatch against a stub completion to confirm what actually goes out: gpt-3.5-turbo.

Expected behavior

Either the default resolves to a model with no published shutdown date, or model becomes required on this path so the failure is a clear argument error at call time rather than a provider error that starts on 2026-10-23.

Libraries used w/ versions:

guardrails-ai==0.11.0 (latest release at time of filing). I read the file at tag v0.11.0 and on main @ 06d0ff2c — the line numbers above are identical in both, so this is in the released wheel and not only on main.

Environment details:

Not environment-specific — this is a static read of the source, and the trigger is the calendar, not the platform.

Additional context

Two things that may widen the fix:

• The async twin has no model parameter at all (llm_providers.py:633-641, AsyncLiteLLMCallable.invoke_llm(self, text=None, instructions=None, messages=None, *args, **kwargs)), and its acompletion(*args, **kwargs) call passes through whatever the caller gave. So for identical caller input with no model=, the sync path substitutes gpt-3.5-turbo and the async path leaves the choice to litellm — the two diverge today, independently of the date.
• The same id is recommended in the project's own docs: the docstrings at llm_providers.py:162 and :650 both show model="gpt-3.5-turbo", as do several notebooks under docs/how_to_guides/ (streaming.ipynb, streaming_structured_data.ipynb, async_streaming.ipynb, remote_validation_inference.ipynb).

One thing I checked and am deliberately not reporting, so it doesn't get swept in: the gpt-4 / gpt-4-0613 occurrences in utils/openai_utils/streaming_utils.py:37-42 are the verbatim OpenAI-cookbook token-counting table. They compute locally via tiktoken, never touch the network, and nothing in the repo imports that module — a retirement cannot break them.

I searched open and closed issues and PRs for prior reports of model retirement before filing and found none. (#1638 / update-hub-retirement-date is about the Guardrails Hub sunset, not OpenAI model deprecations.)

Happy to open a PR for the mechanical part — making model explicit, or raising at construction when it's absent. Which model id should replace the default is a cost and behaviour decision for this project, so I'd leave that choice to you rather than pick one in a patch.

Found with a scanner I run that checks pinned model ids in source against published provider shutdown dates; it flagged this one for human review rather than changing it.