sarvesh-pathak opened a new issue, #77258:
URL: https://github.com/apache/airflow/issues/77258

   ### Under which category would you file this issue?
   
   Providers
   
   ### Apache Airflow version
   
   main (d0cd1700c4728419e04652212c9fd254197cfd9c)
   
   ### What happened and how to reproduce it?
   
   `OpenAIHook` and `OpenAIResponseOperator` default `model` to `gpt-4o-mini`:
   
   - 
https://github.com/apache/airflow/blob/d0cd1700c4728419e04652212c9fd254197cfd9c/providers/openai/src/airflow/providers/openai/hooks/openai.py#L319
 (`create_chat_completion`)
   - 
https://github.com/apache/airflow/blob/d0cd1700c4728419e04652212c9fd254197cfd9c/providers/openai/src/airflow/providers/openai/hooks/openai.py#L331
 (`create_response`)
   - 
https://github.com/apache/airflow/blob/d0cd1700c4728419e04652212c9fd254197cfd9c/providers/openai/src/airflow/providers/openai/hooks/openai.py#L343
 (`parse_response`)
   - 
https://github.com/apache/airflow/blob/d0cd1700c4728419e04652212c9fd254197cfd9c/providers/openai/src/airflow/providers/openai/operators/openai.py#L227
 (`OpenAIResponseOperator`)
   
   OpenAI's deprecations page lists "Jan 14, 2027 | GPT-4o mini 
(`gpt-4o-mini-2024-07-18`) | Deprecated | `gpt-6-luna`" 
(https://developers.openai.com/api/docs/deprecations, read 18 Nov 2026). After 
that date, any task that relies on the default model will fail.
   
   The common.ai example DAGs and connection placeholders use models that shut 
down sooner (same page):
   
   - `openai:gpt-5`: Dec 11, 2026, replacement `gpt-6.1-sol`.
     - `example_dags/example_langchain_hook.py` lines 38, 88 and 116; 
`example_langchain_tool_agent.py:88`; `example_langchain_toolset_bridge.py:49`
     - The connection form placeholder `{"model": "openai:gpt-5"}` 
(`hooks/pydantic_ai.py:301`, `get_provider_info.py:196`) and `"llm_model": 
"gpt-5"` (`hooks/llamaindex.py:164`, `get_provider_info.py:460`)
     - Most common.ai docs pages, starting with `docs/quickstart.rst` and 
`docs/installation.rst`
   - `openai:gpt-5-mini`: Dec 11, 2026, replacement `gpt-6.1-terra`.
     - `example_llm_batch.py` lines 91 and 138; `example_llm_fallback.py` lines 
27 and 29; `example_agent_recovery.py:28`
     - `docs/frameworks/openai_agents.rst` and `docs/retry_policies.rst`
   - `openai:gpt-4.1-mini`: Jan 14, 2027, replacement `gpt-6-luna`. 
`example_llm_batch.py:170`.
   - `anthropic:claude-sonnet-4-5`: Anthropic retires Claude Sonnet 4.5 on 30 
Nov 2026, replacement `claude-sonnet-5-5` 
(https://platform.claude.com/docs/en/models/sonnet-5-5/migration-guide, read 18 
Nov 2026). It appears in the `AirflowDurability` docstring example and in 
`docs/connections/pydantic_ai.rst` (lines 91 and 170):
     
https://github.com/apache/airflow/blob/d0cd1700c4728419e04652212c9fd254197cfd9c/providers/common/ai/src/airflow/providers/common/ai/durable/capability.py#L384
   
   I left out the `azure:gpt-5` placeholders. Those are Azure deployment names 
with their own dates.
   
   **Assistants methods.** The 16 Assistants/Threads hook methods deprecated in 
#69071 still warn that the API "will be removed by OpenAI on 2026-08-26":
   
   
https://github.com/apache/airflow/blob/d0cd1700c4728419e04652212c9fd254197cfd9c/providers/openai/src/airflow/providers/openai/hooks/openai.py#L74-L78
   
   OpenAI now lists the Assistants API under past deprecations, with a shutdown 
date of Aug 26, 2026 and "Responses API and Conversations API" as the 
replacement (same page). So these methods can no longer succeed. 
`create_assistant` also defaults to `gpt-4o-mini` (line 419).
   
   ### What you think should happen instead?
   
   1. Change the four `gpt-4o-mini` defaults to `gpt-6-luna`, and note the 
change in the provider changelog.
   2. Update `providers/openai/docs/operators/openai.rst` lines 136-138. The 
text says the default "is not a reasoning model", but `gpt-6-luna` is one. Its 
model page says "`reasoning.effort` supports `none`, `low`, `medium` (default), 
`high`, `xhigh`, and `max`" 
(https://developers.openai.com/api/docs/models/gpt-6-luna, read 18 Nov 2026). 
The operator's existing note that `output_text` can be empty under a small 
`max_output_tokens` would then apply to the default too.
   3. The same page says "Chat Completions supports function calling only with 
`reasoning_effort` set to `none`." `create_chat_completion` passes `tools` 
through, so a docs line for that case would help.
   4. In the common.ai examples, placeholders and docs, change `gpt-5` to 
`gpt-6.1-sol`, `gpt-5-mini` to `gpt-6.1-terra`, `gpt-4.1-mini` to `gpt-6-luna` 
and `claude-sonnet-4-5` to `claude-sonnet-5-5`. The batch example that sends 
`reasoning_effort: "low"` still works on `gpt-6.1-terra`, which accepts `low`. 
The Sonnet 4.5 examples send no `temperature`, `top_p` or thinking settings, so 
the ID change is enough for those.
   5. For the Assistants methods, put the reason string in the past tense. Then 
either remove the methods in the next major provider release or make them raise 
a clear error.
   
   The hook and operator don't add `temperature` or `top_p` themselves. 
OpenAI's model pages don't say whether `gpt-6-luna` and `gpt-6.1-*` accept 
them, and I haven't run any of these tasks against the new models.
   
   ### Operating System
   
   Not OS-specific
   
   ### Deployment
   
   Other
   
   ### Apache Airflow Provider(s)
   
   openai, common-ai
   
   ### Versions of Apache Airflow Providers
   
   apache-airflow-providers-openai 2.2.0 and apache-airflow-providers-common-ai 
1.1.0 (the versions on main)
   
   ### Anything else?
   
   Happy to open a PR for this if it's useful.
   
   Found with Vitasta, a source-available scanner for AI model and API 
retirements: https://github.com/vitasta-dev/vitasta
   
   ### Are you willing to submit PR?
   
   - [x] Yes I am willing to submit a PR!
   
   ### Code of Conduct
   
   - [x] I agree to follow this project's Code of Conduct
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

  • [I] OpenAI provider: default m... sarvesh-pathak (via GitHub) via commits

Reply via email to