OpenAI shuts Agent Builder, Evals and reusable prompts on 30 November 2026

Summary. OpenAI is retiring three developer-platform surfaces on the same day. Agent Builder, the Evals platform and reusable prompt objects (v1/prompts) all shut down on 30 November 2026, having been announced deprecated on 3 June 2026. That is 180 days of notice, which matches the 6-month minimum OpenAI publishes for generally available models — a policy its deprecations page never extends to products. The announcement was one 40-word entry in the API changelog; there is no first-party blog post. The date most teams will miss is not 30 November. It is 31 October 2026, when existing evals "become read-only" — and OpenAI's own migration cookbook requires a completed run before an eval can be exported as a runnable config. Miss that window and the last 30 days are salvage, not migration.

What OpenAI actually announced, and where

The whole of the first-party announcement reads, verbatim, in the changelog entry dated Jun 3:

"Announced the deprecation of reusable prompt objects, the Evals platform, and Agent Builder. See the deprecations page for shutdown timelines and migration guidance."

The deprecations page carries the dates. Three separate sections, three tables, one shutdown date:

Surface	Announced	Interim milestone	Shutdown	Named replacement
Agent Builder	3 June 2026	none stated	30 November 2026	Agents SDK or ChatGPT Workspace Agents
Evals platform	3 June 2026	31 October 2026, "existing evals become read-only"	30 November 2026	Promptfoo (third party)
Reusable prompts (v1/prompts)	3 June 2026	3 June 2026, prompt creation "de-emphasized"	30 November 2026	prompt content in application code
Graders	3 June 2026 (implied)	none stated	not stated	Promptfoo assertions
Self-serve fine-tuning	7 May 2026	2 July 2026	6 January 2027	none stated

The graders row is not an oversight on our part. The graders guide says OpenAI "is deprecating graders as part of the evals and fine-tuning workflows they support," and the deprecations page punts on the date: "Fine-tuning-related timelines remain covered in the self-serve fine-tuning section below." That section gives 6 January 2027. A team using graders inside fine-tuning therefore has two candidate deadlines 37 days apart and no page that reconciles them.

The 31 October trap

This is the part worth acting on this week.

OpenAI's recommended replacement for Evals is Promptfoo, an open-source CLI described in the migration cookbook as "an open-source CLI and library for evaluating and red-teaming AI apps and agents." Its Community tier is $0 under an open-source licence and includes up to 10,000 red-team probes per month, per Promptfoo's pricing page; Enterprise is quote-only. So the migration cost here is engineering time, not licence fees.

The export procedure in the cookbook contains this instruction: "Select a completed run to use as the basis for the exported configuration. If your evaluation does not have a completed run yet, run it once before exporting."

Now read that against the deprecation table: from 31 October 2026, "Existing evals become read-only." You cannot run a read-only eval. Any eval that has not completed a run by 31 October cannot be exported as a runnable Promptfoo config at all — it falls into the cookbook's manual-recreation path instead. The month between 31 October and 30 November is not a grace period. It is a window for downloading results you can no longer regenerate.

Three more things do not survive the export, quoted from the cookbook's own table of cases requiring manual setup: "A grader is not included in the exported configuration." "A tool, agent, or custom provider workflow is not reproduced." And on scoring: "Similarity-based scoring may not produce identical numerical results across systems. Any manually recreated grader, especially an LLM-as-a-judge grader, should be validated before you rely on it for regression decisions."

Run history is a separate artefact again. You download results, then promptfoo import <downloaded-results-file>, which "preserves past run results for reference. It does not create the runnable Promptfoo configuration used for future evaluation runs." Nothing on any OpenAI page commits to retaining eval history past 30 November. If your release gate cites a numeric score from a historical run, export the file. This is the same class of failure we described in AI agent evals that fail silently in CI/CD: the eval keeps returning a number long after the number stopped meaning anything.

Agent Builder: an export, not a converter

The Agent Builder migration guide is blunt about what it does and does not do: "Use this guide to export an existing Agent Builder workflow as Agents SDK code," followed immediately by "This process does not convert your workflow graph or guarantee that every behavior transfers unchanged."

The mechanics are manual. Open the workflow, select Code in the top navigation, select Agents SDK, pick TypeScript or Python, copy the export. Two destinations are offered. The Agents SDK path is "Best for building agents through code." The ChatGPT Workspace Agents path is "Best for building agents through natural language," requires "a ChatGPT Business, Enterprise, or Edu workspace with access to workspace agents and permission to create agents," and the documented procedure is to paste your exported code into a chat box under the prompt "Please help me convert this workflow into an agent."

The stated limits matter more than the steps. "Workflows with strong determinism at their core may not migrate faithfully to a workspace agent." "Connected apps, authentication, publishing, and permission configuration require separate review in ChatGPT." "An Agents SDK implementation requires you to validate your application's runtime configuration, tools, authentication, permissions, and deployment."

Read plainly: authentication, connected apps and permissions do not migrate. Those are the parts that take the longest and the parts a security review will stop on.

Three OpenAI pages, three different answers

The most useful finding is not a date. It is that OpenAI's documentation does not agree with itself about where an Agent Builder user should go.

The ChatKit guide says ChatKit survives — "OpenAI is deprecating Agent Builder... ChatKit is still available" — and then demotes the hosted route: "Use this path only if you already have an Agent Builder workflow that backs your ChatKit implementation." The surviving mode is the one where you run your own server-side agent.

The Agent Builder overview page carries the deprecation banner and, further down the same page, still recommends the route the ChatKit page just warned against: "ChatKit: Follow the ChatKit quickstart and pass in your workflow ID to embed this workflow into your application. If you're not sure, we recommend this option."

The official migration guide, meanwhile, offers ChatKit as neither of its two options and does not mention it once.

Because a ChatKit session is created by passing a workflow ID, any live ChatKit embed backed by a hosted Agent Builder workflow should be treated as breaking on 30 November 2026. No OpenAI page states that published workflows keep running. The strongest language anywhere is "Existing users can continue using it during the transition window."

Two more banner gaps are worth a grep of your own bookmarks. The agent-evals guide carries no deprecation notice at all and actively routes readers into the dying product: "If you need advanced features such as evaluation against external models, evaluation APIs, or larger-scale batch evaluation, use Evals alongside datasets." And the Evals guide suggests Datasets as a gentler starting point — a page that carries the identical shutdown banner.

Reusable prompts: the cache regression nobody flags

The prompt-object migration guide states the timeline directly: "Prompt creation will be de-emphasized beginning June 3, 2026, and v1/prompts is scheduled to shut down on November 30, 2026."

The shape of the change is that prompt IDs, server-side versions and the variables map stop existing as API concepts. You send model plus an input array of messages. OpenAI's own list of what to change includes "Move versioning to your repo using git commits, PR review, and tests or evals" — which lands a fortnight before the eval tooling you would use for that is switched off.

The buried cost is the last bullet: "Keep static content first and dynamic content later to preserve prompt caching benefits, since cache hits depend on exact prefix matches."

A prompt object let you interpolate variables wherever they read best. Application code lets you do the same thing, and if you interpolate a customer name near the top of a long system prompt, every request becomes a cache miss. Nothing errors. The bill just goes up. Teams that ported prompts mechanically during the 23 July 2026 model shutdowns hit the same class of silent regression.

India-specific considerations

For teams in India running on OpenAI's platform, two points compound. First, the surviving eval path is a third-party tool, which means eval inputs — often real user prompts — move to a new processor. Under the Digital Personal Data Protection Act 2023, that is a change of processor for any eval dataset carrying personal data, and it needs to appear in your records before 30 November, not after. Promptfoo's Community tier runs locally and its Enterprise on-prem option deploys inside your own cloud account, so keeping the data in-country is achievable, but it is a deployment decision someone has to make.

Second, the read-only date falls on 31 October 2026, in the middle of the Indian festive-release freeze that most product teams observe. If your change-freeze runs through Diwali, the practical export deadline is earlier still.

What to do in the next two weeks

Inventory first, in this order, because the deadlines are not the same.

Run every eval you intend to keep, once, before 31 October 2026, then export the runnable Promptfoo config. Do this even for evals you are unsure about; a completed run costs an API call and buys you the export.

List every ChatKit embed and check whether it passes a workflow ID. If it does, it is on the 30 November clock.

Grep your codebase for pmpt_ prefixes and prompt= parameters on Responses calls. Move that content into source files, and put the static portion first.

Recreate graders as Promptfoo assertions and validate them against a historical run before you trust either. LLM-as-a-judge graders in particular need a side-by-side check.

If you fine-tune, treat 6 January 2027 as a separate deadline and confirm which graders your jobs depend on.

The real cost is usually the migration, not the code. Teams that treat this as a find-and-replace tend to discover the authentication and permissions work only after the export lands.

What is still unknown

OpenAI has not published a retention policy for eval run history after 30 November 2026. It has not stated whether Agent Builder workflows already published continue serving. It has not given graders a shutdown date. And it has not issued a blog post for any of it — three developer products were retired in a 40-word changelog line, which is itself the signal about how much notice similar surfaces should be assumed to get.

FAQ

When exactly does OpenAI Agent Builder shut down?

Agent Builder shuts down on 30 November 2026, according to OpenAI's deprecations page. The deprecation was announced on 3 June 2026, giving 180 days of notice. No OpenAI page states that already-published workflows keep serving traffic after the shutdown date, so treat live embeds as breaking.

Why does 31 October 2026 matter more than 30 November?

From 31 October 2026, existing OpenAI evals become read-only. The migration cookbook requires a completed run before an eval can be exported as a runnable Promptfoo config. An eval with no completed run by that date cannot be exported as a runnable config, leaving only manual recreation in the final month.

What replaces the OpenAI Evals platform?

OpenAI recommends Promptfoo, an open-source command-line tool and library for evaluating and red-teaming AI applications. Its Community tier is free and includes up to 10,000 red-team probes per month; Enterprise pricing is quote-only. Graders, custom providers and agent workflows are not carried across by the export.

Does ChatKit survive the Agent Builder shutdown?

Yes. The ChatKit guide states ChatKit is still available. However, the hosted path that passes an Agent Builder workflow ID is explicitly demoted, and OpenAI recommends running your own server-side agent implementation instead. Any ChatKit embed backed by a hosted workflow should be treated as breaking on 30 November 2026.

What happens to reusable prompt objects and prompt IDs?

The v1/prompts API shuts down on 30 November 2026. Prompt IDs, server-side prompt versions and the variables map stop existing as API concepts. Prompt content moves into application source code and is passed through the input array on Responses API calls instead, with versioning handled in your repository.

Will moving prompts into code raise our token bill?

It can. OpenAI advises keeping static content first and dynamic content later, because prompt cache hits depend on exact prefix matches. Interpolating a variable near the top of a long system prompt turns every request into a cache miss. Nothing errors when this happens, so the regression shows up only on the invoice.

Do OpenAI's own documentation pages agree on the migration path?

No. The Agent Builder overview recommends embedding via ChatKit with a workflow ID. The ChatKit guide says to use that path only if you already have such a workflow. The official migration guide offers only the Agents SDK and ChatGPT Workspace Agents, and does not mention ChatKit at all.

When do graders stop working?

OpenAI has not published a shutdown date for graders. The graders guide says they are being deprecated as part of the evals and fine-tuning workflows they support. The Evals surfaces end on 30 November 2026, while self-serve fine-tuning job creation ends on 6 January 2027, leaving two candidate dates 37 days apart.

How eCorpIT can help

We run this kind of platform migration as a scoped exercise rather than a code sweep: inventory the evals, prompt objects and workflow embeds that are actually load-bearing, export what can still be exported before the read-only date, and rebuild the graders and CI gates that do not carry across. Our AI evals and observability service covers the rebuild and the regression checks around it, and we design the data handling aligned with DPDP Act 2023 requirements when eval datasets move to a new processor. eCorpIT is CMMI Level 5, MSME Certified and ISO 27001:2022 certified. Talk to our senior engineering team via /contact-us/ if you need the 31 October window covered.

References

1. OpenAI API deprecations — shutdown tables for Agent Builder, Evals, reusable prompts and self-serve fine-tuning.
2. OpenAI API changelog — the 3 June 2026 deprecation entry.
3. Migrate from Agent Builder — export procedure and stated migration limits.
4. Migrate from prompt objects — v1/prompts timeline and prompt-caching guidance.
5. Moving from OpenAI Evals to Promptfoo — export requirements and cases needing manual setup.
6. OpenAI ChatKit guide — ChatKit status and the demoted hosted path.
7. OpenAI graders guide — graders deprecation wording.
8. OpenAI agent evals guide — routes readers to Evals with no deprecation banner.
9. OpenAI evals guide — recommends Datasets, which carries the same shutdown banner.
10. Promptfoo pricing — Community, Team and Enterprise tiers.
11. Promptfoo — open-source evaluation and red-teaming tool.
12. Ministry of Electronics and IT, Digital Personal Data Protection Act 2023 — processor obligations referenced in the India section.

Last updated: 20 August 2026.