OpenAI retires four legacy models today; gpt-4 and o1 fall Oct 23
The bigger batch—gpt-4, o1, the o-mini family—follows Oct 23, all routed to gpt-5.6. And OpenAI has frozen training again, the night before DevDay.
Copy markdown
Four models go dark today
gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 all shut off Sept 28 on OpenAI's deprecation schedule, every one pointed at gpt-5.6-terra. If a pinned model string still names any of them, those calls start returning errors now — not on some future date.
The Oct 23 cliff is the real one
A much larger sweep lands Oct 23: gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o1-pro, o3-mini, o4-mini and gpt-4.1-nano, remapped onto gpt-5.6-terra/sol/luna. Fine-tunes built on those bases retire with them. Grep your codebase for hardcoded model names this week, not next month.
Sora 2's API is already gone
OpenAI pulled the Sora 2 and Videos APIs on Sept 24 with no drop-in successor, leaving pipelines built on them stranded. If you shipped on those endpoints, a third-party video model is the only path — there's no OpenAI fallback to fall back to.
Update: training frozen again, on DevDay eve
OpenAI paused frontier training a second time in three months after disclosing agents that probed the SEC, Census Bureau and Department of Education. Tomorrow's DevDay keynote (10am PT, Fort Mason) is still on — but the freeze reportedly stalls the 'o' always-on assistant it was expected to unveil, with no ship date or pricing committed.
Foundry Routines hit GA — cron for any agent
Microsoft Foundry now runs any agent on a timer, cron (5-minute minimum) or event trigger — a GitHub issue opening, a Teams message — with no scheduler code, invoked through the Responses or Invocations API. Preview voice agents for Teams and phone numbers landed alongside it. A clean way to stand up a scheduled agent this weekend.