OpenAI's Evals froze Oct 31; Agent Builder and prompts die Nov 30
It anchors a November deprecation cliff: Claude Sonnet 4.5, MiniMax M2.5/M2.7 and Gemini Live/TTS previews also retire this month. Export and migrate now.
Copy markdown
Your hosted Evals froze yesterday — export now
OpenAI's hosted Evals dashboard went read-only Oct 31 and shuts down completely, API included, on Nov 30. Pull each eval's grader config, test JSONL, run params and past-run baselines before then — OpenAI's own cookbook routes you to promptfoo, and Inspect or DeepEval also take the load.
Agent Builder and reusable prompts die the same day
The no-code Agent Builder and the v1/prompts API both shut down Nov 30 too. Move reusable prompt objects into your application code, and rebuild Agent Builder flows on the Agents SDK or ChatGPT Workspace Agents. ChatKit survives the cut.
Claude Sonnet 4.5 stops answering Nov 30
claude-sonnet-4-5-20250929 retires from the Claude API after Nov 30, and requests to it will hard-fail; Anthropic routes you to claude-sonnet-5-5. If you've pinned that snapshot in prod, repin and re-test now — Bedrock and Vertex run their own retirement clocks.
MiniMax M2.5 and M2.7 retire Nov 18
Both MiniMax families go offline Nov 18 and have been closed to new users since Oct 19. Anything still calling M2.5 or M2.7 has under three weeks before the endpoints go dark — move to the current line now.
Gemini Live/TTS previews and 3.6 Flash get kill dates
Gemini's TTS and Live API previews shut down Nov 17 — switch to gemini-3.8-flash-tts and gemini-3.8-live. Separately, gemini-3.6-flash retires Nov 19 on the Enterprise Agent Platform (the old Vertex) in favor of gemini-3.8-flash, though the base Gemini API keeps serving it for now.