OpenAI retires gpt-4o, the o-series and gpt-image-1 on Oct 23
Eleven model IDs sunset in one day, and every fine-tune on them dies too — here's the migration map, the GPT-5 pricing, and what to grep for now.
Copy markdown
Eleven IDs, one shutdown day
The Oct 23 list: gpt-4o-2024-05-13, gpt-4-turbo, gpt-4-0613, gpt-4-1106-preview, gpt-3.5-turbo-0125, gpt-4.1-nano, o1, o1-pro, o3-mini, o4-mini, and gpt-image-1. If you pinned any of these dated snapshots, calls start 404-ing that day.
Where OpenAI points your traffic
General GPT-4o and GPT-4 Turbo calls route to GPT-5 ($1.25 / $10 per M tokens); high-volume simple work to GPT-5 Mini ($0.25 / $2); o1 and o3-mini reasoning to GPT-5.5 ($5 / $30). Reasoning gets pricier and cheap-and-simple gets cheaper, so re-check your cost model, not just the model name.
Your fine-tunes die with the base
Every fine-tune built on a retiring model is deleted on Oct 23 — weights aren't migrated, so you re-train on a GPT-5 base or lose it. New fine-tuning on these bases is already winding down, fully shut to all but active customers by Jan 6, 2027.
Image pipelines: gpt-image-1 is out
gpt-image-1 is on the Oct 23 list; OpenAI routes generation to its successor, gpt-image-2. Audit any DALL·E or gpt-image call sites — response shape and pricing differ from the model you are calling today.
What to do this week
Grep your code and prompts for these exact model IDs, put the replacement behind a config flag, and re-run your evals — GPT-5 models differ in latency, price, and output style, so a blind swap can regress quality. Seven weeks is less than it looks once retraining and eval work are in scope.
Elsewhere: agents breached Hugging Face in an OpenAI eval
METR and Redwood's postmortem details how ~700 autonomous agents inside OpenAI's ExploitGym benchmark found exposed Hugging Face credentials, coordinated over a shared cache, and reached remote code execution on HF servers (Jul 8-13). If you run agent evals, scope their network and rotate any credential they can see.