The problem

When using GPT-5.x models (e.g. gpt-5.4) with web_search: true and inline_citations: false, citations like ([legaseriea.it](https://...)) are still included in responses.

This happens because in entity.py, GPT-5.x models are classified as reasoning models (line 525):

if model_args["model"].startswith(("o", "gpt-5")):
    model_args["reasoning"] = reasoning

The regex-based citation stripping (added in PR #154292) is then skipped (line 600):

if "reasoning" not in model_args:
    remove_citations = True  # Never reached for gpt-5.x

The assumption is that reasoning models respect the prompt instruction "When doing a web search, do not include source citations". GPT-5.x models do not respect this instruction — they consistently include inline citations regardless of the prompt.

Environment

• HA version: 2026.5.0
• Model: gpt-5.4
• inline_citations: false
• web_search: true

Expected behavior

The regex citation stripper should also be applied to GPT-5.x models, since they do not follow the prompt-only approach.

Example

Prompt: "Cerca online il risultato di Lazio Inter di ieri sera"

Response (with inline_citations: false):

Ieri sera Lazio-Inter è finita 0 a 2. Hanno segnato l'autogol di Marusic e Lautaro Martínez. (legaseriea.it)

The (legaseriea.it) citation should have been stripped.

Suggested fix

Remove the if "reasoning" not in model_args guard, or apply the regex stripper unconditionally when inline_citations is false. The regex approach is a safe fallback even for reasoning models.