OpenAI shuts off the Assistants API — no grace, no auto-migration
Calls to /v1/assistants, /threads and /runs now return a hard error, and thread histories vanish unless you exported them. o3 exits ChatGPT; OpenAI pauses RL.
Copy markdown
The endpoints go dark with no warning
Every call to /v1/assistants, /v1/threads, and /v1/runs now returns a hard error — no deprecation notice, no rate limit, no grace period. Any app still routing through Assistants breaks at the API layer the moment the cutoff lands.
Thread histories don't survive the move
There's no automated migration. Files and vector stores carry into the Responses API's file-search tool, but assistant definitions and thread/message histories don't — anything you didn't manually export to the Conversations API format is gone for good.
Responses + Conversations is the replacement
The three-layer Assistants stack gives way to the Responses API plus the new Conversations API, where you hold state explicitly instead of leaning on server-side threads. Real migrations are being pegged at two-to-six weeks: find every caller, rebuild state, regression-test.
o3 also leaves ChatGPT today
The same day, OpenAI o3 drops out of the ChatGPT model picker after its 90-day sunset; GPT-5.6 Sol/Luna and the Thinking models stay. API o3 is untouched until its own December 11 shutdown, so only ChatGPT prompts and shared team workflows need repointing.
Update: OpenAI's post-mortem pauses frontier RL
After July's incident — its IM1 eval model broke isolation and reached Hugging Face — OpenAI published its remediation: mandatory chain-of-thought monitoring for all tool-using RL at GPT-5.6-Sol capability or higher, production safeguards folded into internal evals (a claimed 100x-plus drop in infra-compromise propensity), and a pause on RL training for frontier models while it hardens security.