OpenAI 停用 GPT-4o、4.1 和 5-nano：如何找出程式碼庫中的所有模型 ID

OpenAI's deprecations page has a busy few months coming up. Some model IDs already stopped working on Oct 23, 2026. More shut down on Dec 11, then Jan 14, then Apr 1. If a model string is hard-coded in a config file, a .env, or a test fixture, you won't find out until requests start failing.

Here are the dates that matter (checked against the page on Oct 31, 2026) and a free one-liner that finds them.

The dates

This is a summary. The page lists dated snapshots for most rows, so check it for the exact ID you use. OpenAI also notes that fine-tuned models follow their base model's schedule.

One catch: the bare aliases gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini aren't rows on the page. They point at snapshots that are on it, so treat them as Jan 14 and confirm which snapshot your account resolves them to.

The one-liner

Run this from your repo root. It needs GNU grep (for -P; on macOS, brew install grep and use ggrep):

grep -rnIoP --exclude-dir={.git,node_modules,venv,.venv,dist,build} \
  '(?<![\w.-])(gpt-4o(-mini)?(-(2024-[0-9]{2}-[0-9]{2}|tts|realtime|audio|transcribe))?|gpt-4\.1(-mini|-nano)?(-2025-04-14)?|gpt-5(-mini|-nano|-pro)?(-20[0-9]{2}-[0-9]{2}-[0-9]{2})?|gpt-5\.4-nano|gpt-5\.1|o1|o3(-mini|-pro)?|o4-mini|gpt-3\.5-turbo|gpt-4-turbo|tts-1(-hd)?|whisper-1)(?![\w.-])' . \
  | sort | uniq -c | sort -rn

It covers .env, YAML and JSON as well as code, because -r reads every text file. The lookarounds stop it matching gpt-5 inside gpt-5.6-luna. You get count path:line:match, which is enough to open a ticket per service.

Watch out for false positives. o1 and o3 are short, so the regex can catch things that aren't model names. Read the matches before you change anything.

What grep can't see

• IDs built at runtime, like f"gpt-{version}". Search for "gpt-" string fragments too.
• IDs stored in a database or a feature flag service. Query those directly.
• Third-party SDK defaults. A library may pick a model for you. Check its config.
• Other repos. Infra-as-code, notebooks and Zapier-style automations often have their own copies.

A safe swap order

1. Fix anything that's already broken first. Those calls are failing now.
2. Change the model ID in one place (an env var or a config constant), not across 40 files.
3. Run your evals or a small golden set on the replacement before you switch traffic. Newer models can behave differently on structured output and tool calls.
4. Add a CI check so the old IDs don't come back in a copy-pasted snippet.

I packaged this into a small scanner that reads the dates from a JSON file, marks rows past their date as BROKEN NOW, writes a CSV/markdown swap list and can fail a CI job. If that saves you time, it's here. The grep above works fine on its own.