Picture a small support team that built its reply assistant on OpenAI’s platform. The system prompt is saved as a reusable prompt object. The routing — triage, look up the order, draft a reply, escalate refunds — is a workflow drawn in a visual builder. And forty tricky customer emails sit in an eval set, so every prompt change gets tested against the cases that broke it last time.
On November 30, 2026, all three of those containers close on the same day.
What changes on November 30
OpenAI’s deprecations page lists three related shutdowns, all announced on June 3, 2026:
| What | Status on the page | OpenAI’s recommended replacement |
|---|---|---|
/v1/prompts and reusable prompt objects |
Shuts down Nov 30, 2026 | “Move prompt content into your application code.” |
| Agent Builder | Shuts down Nov 30, 2026 | “Agents SDK or ChatGPT Workspace Agents.” ChatKit remains available. |
| Evals dashboard, API, and graders for eval workflows | Already read-only; Nov 30, 2026 is the listed date | “Migrate to Promptfoo” |
Separately, Anthropic’s model deprecations page retires claude-sonnet-4-5-20250929 on the Claude API on the same day, with claude-sonnet-5-5 as the replacement. If you run Claude through Amazon Bedrock or Google Vertex AI, check those platforms’ own dates; Anthropic’s page notes they are set separately.
None of this is careless. The OpenAI notices went out on June 3, nearly six months ahead. Clear notice still leaves the same question: where did your work actually live?
OpenAI’s own advice is to move it into files
OpenAI’s prompt-object migration guide tells developers to “store the prompt text in your codebase” and pass it to the model directly, with versioning handled in your own source control.
That is sound advice, and it matches what we argued for assistant instructions in Own Your AI Instructions. The November 30 changes extend the same point to two things that are easier to forget: the workflow around the prompt, and the cases that prove it works.
A console holds three kinds of context, not one
| What lived in the console | What it really is | Where it belongs |
|---|---|---|
| Saved prompt | The instruction: role, rules, tone, output format | A plain instruction file, versioned, with a short changelog |
| Visual workflow | The process: steps, branches, handoff points, when a human steps in | A written process note — step list, decision rules, escalation conditions — kept next to the code that runs it |
| Eval set | The evidence: the cases that define “working” | A folder of test cases with expected behaviour, in plain text or a simple table |
Most migration guidance stops at the first row. The second and third rows are where most of the effort went.
The eval cases are the clearest example. Forty hard emails, each with a note on what a good reply must and must not do, record where the assistant has failed before. That knowledge is not tied to any one evaluation tool. If it exists only inside one, it becomes hard to reach when the tool changes — and the next prompt change ships untested.
Exported code is not a migrated workflow
OpenAI’s Agent Builder migration guide does give an export path: open the workflow, select Code, choose Agents SDK, and copy the TypeScript or Python. Use it. But read its own caveat: the process “does not convert your workflow graph or guarantee that every behavior transfers unchanged.”
Generated code tells you what the workflow did. It does not tell the next person, or the next agent, why it branches where it does, which step is allowed to send anything, or when a refund must go to a human. That reasoning was in the diagram and in your head. Write it down once, in words, and the next migration is a re-implementation rather than an excavation.
The prompt-object guide, by contrast, does not describe an export. Its instruction is to move the prompt content into your code. If you have prompts saved only on the platform, copy them out before November 30 rather than after.
What this does not solve
Moving prompts, process and test cases into files does not port them to a new model. The same instruction can behave differently on a replacement model, which is exactly why the eval cases matter. Files do not run tests for you; you still need a test runner, whether that is Promptfoo, a script, or a careful manual pass. And a file nobody maintains drifts just as surely as a console nobody opens.
What files give you is a recovery point you own. When a feature closes, a model retires or you change tools, you re-run known cases against a known instruction instead of reconstructing both from memory.
Try it before November 30
Set aside thirty minutes for one assistant or agent you depend on.
- List what lives only in a vendor console. Saved prompts, workflow diagrams, eval sets, model IDs pinned in settings. If it would vanish with the account, it goes on the list.
- Copy every saved prompt into a plain file. One file per assistant. Add a line at the top with today’s date and the model it was tuned on.
- Write the workflow as steps. Number them. For each, note what it reads, what it may write or send, and the condition that hands it to a human.
- Copy out the test cases. For each case, keep the input and one or two lines on what a good answer must and must not include. If your evals live in OpenAI Evals, plan for this to be manual: OpenAI’s Promptfoo migration guide covers “manually recreating an evaluation” and does not rely on an export feature.
- Check your model IDs. Search your code and settings for any model with a retirement date on the vendor’s deprecations page, and record the replacement next to it.
- Test the result. Hand the files — and only the files — to a fresh session or a colleague and ask them to describe what the assistant does and how they would know it still works. If they cannot, the files are not finished.
Nothing in this exercise is specific to one vendor. It is the same work for any tool that offers to hold your instructions for you.
Keep the parts that took the longest
Prompts get the attention, but the expensive parts of an AI workflow are the process you designed around it and the cases you collected to prove it works. Both outlive any one console if you keep them as files.
For a structure that keeps instructions, project context and working notes together in files you own, explore the KOS Starter Kit.