The OpenAI Agent Builder shutdown is set for November 30, 2026, and it takes the Evals dashboard and API and the v1 prompts API down with it. If you built anything on those, you have about 18 days to move it.
Nothing here is a rumor. OpenAI posted the dates on its official deprecations page, and it first notified developers on June 3, 2026.
The short version:
- Agent Builder shuts down on Nov. 30, 2026.
- The Evals platform went read-only on Oct. 31 and shuts down Nov. 30.
- The
v1/promptsAPI and reusable prompt objects also end Nov. 30. - OpenAI points to the Agents SDK, Promptfoo and in-code prompts as replacements.
Those four lines are the whole deadline. The rest of this post is about what to do before it hits.
What the OpenAI Agent Builder shutdown actually covers
Three separate products share one date, which is why this is easy to under-react to. You might use only one of them and miss that the others are going too.
| Product | Read-only | Shuts down | OpenAI’s suggested replacement |
|---|---|---|---|
| Agent Builder | No read-only date listed | Nov. 30, 2026 | Agents SDK or ChatGPT Workspace Agents |
| Evals dashboard and API | Oct. 31, 2026 | Nov. 30, 2026 | Promptfoo |
| v1 prompts and prompt objects | No read-only date listed | Nov. 30, 2026 | Prompt text moved into your own code |
The table comes straight from OpenAI’s page. The thing to notice is that the Evals platform is already locked: you can look at old runs but you can’t add new ones.
Evals are already read-only, so start there
If your team used the Evals API for regression checks, those checks have been unable to write new results since October 31. That is the most urgent item on the list, because quality tracking quietly stops before anyone notices.
OpenAI recommends moving to Promptfoo, an open source evaluation tool, and has published a cookbook guide for the switch. The practical first step is to pull out your test cases and graders now, while the dashboard still loads.
I couldn’t find an official export tool on the deprecations page. Treat that as a gap to plan around, not as proof that none exists.
Saved prompts need a new home in your own code
The v1 prompts API let you store a prompt with an ID and a version, then call it with a map of variables. That pattern ends on Nov. 30, so any request that sends a prompt object will fail afterward.
OpenAI’s migration guide says to move the text into your codebase and pass it to the Responses API as input. Here is the order it recommends:
- Copy each stored prompt into a
prompts/module, one named function per prompt. - Turn prompt variables into typed function arguments.
- Build the system and user messages inside that function.
- Keep versions in git, with tests or evals alongside.
- Put static text first and dynamic values last, so prompt caching still works.
That last step is easy to skip and it matters. Caching depends on exact prefix matches, so a variable near the top of your prompt can quietly raise your bill.
This is also a good moment to audit your other OpenAI dependencies. We covered a similar deadline in OpenAI Models in Cursor End Nov 12, and the lesson is the same: hosted conveniences have expiry dates.
Agent Builder workflows: SDK, ChatKit or something else
Agent Builder is the hardest of the three to move, because a visual workflow doesn’t map neatly to a text file. OpenAI’s page names the Agents SDK or ChatGPT Workspace Agents as the destinations, and says ChatKit remains available.
There is one catch worth checking. A thread on the OpenAI developer community says the ChatKit setup that relied on OpenAI hosting your backend from Agent Builder is effectively dead, leaving only the self-hosted path. That is a community claim, not an official statement, so read the current ChatKit docs before you plan around either option.
The same thread says exporting workflows is only possible while the dashboard is still up. Again, unconfirmed, but it is a good reason to export early. If you are weighing how to manage agent tooling afterward, our explainer on Microsoft APM covers one approach.
A realistic plan for the next 18 days
Don’t try to do all three migrations at once. Sequence them by how fast each one hurts you when it breaks.
Week one is for inventory. List every place your code or your team touches Agent Builder, the Evals API or a stored prompt ID, including scripts nobody has opened in months. Search your repositories for v1/prompts and for any request that sends a prompt object with an id.
Week two is for the moves themselves: prompts into code, eval cases into Promptfoo, and workflows into the Agents SDK. Run the new versions next to the old ones for a few days so you can compare outputs before the old path disappears.
Leave a few days at the end for surprises. Deadlines like this one have a habit of exposing a forgotten cron job on the final afternoon.
Frequently Asked Questions
When does the OpenAI Agent Builder shutdown happen?
November 30, 2026. OpenAI told developers on June 3, 2026 that the product was being deprecated.
Can I still use the Evals API?
You can read existing evals, but the platform went read-only on October 31. It shuts down fully on November 30.
Do I need to rewrite my prompts?
If you call them through the v1/prompts API or prompt objects, yes. Move the text into your code and send it as input to the Responses API.
Is there an official way to export my Agent Builder workflows?
OpenAI’s deprecations page doesn’t describe one. Community posts mention a code export, but I couldn’t confirm it in official documentation.
My take
Spend an hour this week, not on Nov. 29. Copy your prompts into code first, since it takes the least effort, then export eval cases, and leave Agent Builder workflows for last because they need real design decisions. If any of this runs in production, assume there is no safety net after the 30th.
Leave a Reply