OpenAI's shutdown calendar: Evals froze today, Agent Builder goes November 30, and voice follows in January
Read in one sitting, the deprecations page is a four-month countdown for anyone who built agents on OpenAI's earlier tools. Support confirmed today that the newest voice model still lacks an event many developers rely on.
Photo: Developer working with multiple screens and laptop. // GaaS News- Existing OpenAI evals became read-only on October 31. The Evals platform, Agent Builder and reusable prompt objects are scheduled to shut down on November 30.
- Text-to-speech models end January 6, the original Realtime and audio models January 20, and Whisper and the transcription models February 26.
- OpenAI's guide says GPT-Live "has no corresponding event marking the end of each spoken response," and support confirmed that today.
OpenAI's shutdown calendar reached its first deadline today. Under the company's deprecations page, existing evals became read-only on October 31, and the Evals dashboard and API, Agent Builder, and the v1/prompts API with its reusable prompt objects are all scheduled to shut down on November 30. Text-to-speech, Realtime and transcription retirements follow by the end of February.
None of those dates is new on its own, and each has its own notice. What the page shows when read top to bottom is different from what any single notice shows: within four months, much of the tooling a small team would have used to build, test and voice an agent on OpenAI goes away. The Assistants API's removal date was August 26. We could not find a major outlet that has laid the dates side by side, so here they are.
The OpenAI shutdown dates
The OpenAI shutdown dates below come from the deprecations page as it read on October 31.
| Date | What ends | OpenAI's suggested path |
|---|---|---|
| Oct 31, 2026 | Existing evals become read-only | Migration guide to Promptfoo |
| Nov 30, 2026 | Evals dashboard and API | Migration guide to Promptfoo |
| Nov 30, 2026 | Agent Builder | Agents SDK or ChatGPT Workspace Agents |
| Nov 30, 2026 | v1/prompts API and reusable prompt objects | Move prompts into application code |
| Jan 6, 2027 | tts-1, tts-1-hd, gpt-4o-mini-tts | gpt-realtime-2.1-mini |
| Jan 20, 2027 | gpt-realtime, gpt-audio and older audio models | gpt-realtime-2.1, gpt-audio-1.5 |
| Feb 26, 2027 | whisper-1, gpt-4o-transcribe and siblings | gpt-live-transcribe or gpt-transcribe |
ChatKit stays, the page notes. Everything else in the first four rows needs a migration inside a month.
Why the OpenAI voice shutdown is the hard one
The OpenAI voice shutdown is the hard one because the suggested replacements are not drop-in swaps. We reported in September that developers had asked OpenAI not to retire gpt-realtime before a true replacement existed, and that support could not "promise an extension or fix timeline." The date has not moved. The page maps gpt-realtime to gpt-realtime-2.1. Builders who move to GPT-Live, OpenAI's newer voice model, meet a gap the migration guide states plainly: "GPT-Live has no corresponding event marking the end of each spoken response."
A developer raised that on the OpenAI forum on September 15, asking how to chain a voice turn, a tool call and another voice turn without much latency. OpenAI's support account answered today: "The migration guide confirms that GPT-Live currently has no equivalent to Realtime's end-of-response events." The documented approach is to delegate tools to the backend and to track playback in the client. For a phone agent, that event decides when the model has stopped talking, so its absence is a design change rather than a rename.
OpenAI text-to-speech has no exact replacement
OpenAI text-to-speech developers have a narrower problem. The page recommends gpt-realtime-2.1-mini for tts-1, which moves a simple HTTP call onto the Realtime API. When a developer pushed back in a forum thread, support replied on October 8 that gpt-audio-1.5 works over a single HTTP request but "still generates a response to a prompt," that "the docs don't guarantee verbatim reading," and that "the retirement notice doesn't name a drop-in replacement for /v1/audio/speech." For anyone reading fixed scripts aloud, that matters. Compliance disclosures on a phone line often have to be said word for word.
What the OpenAI calendar does not show
What the OpenAI calendar does not show is how many developers are affected. The page gives no usage numbers, and OpenAI describes the November dates as scheduled. Retiring old products is also normal. AWS closed its original Bedrock Agents to new accounts this summer, though it says there is no end-of-life date, a softer path than a hard shutdown. Vendors move fast in this market, and many of these replacements are improvements. The cost is the migration work, and it lands on the people who shipped early.
What OpenAI builders should do before November 30
OpenAI builders have 30 days on the first batch, so start with an inventory:
- List every workflow built in Agent Builder and every saved prompt object, and export them.
- Export eval datasets and results while they are still readable.
- Test voice agents on gpt-realtime-2.1 before committing to GPT-Live, since the two differ on turn events.
- If you buy an agent built on OpenAI, ask the vendor which of these dates touch your deployment.
