From OpenAI's deprecations page and its migration, ChatKit and Evals pages · Updated

OpenAI Agent Builder, prompt objects and Evals shut down on 30 November 2026

OpenAI's deprecations page lists three features with a shutdown date of 30 November 2026, all announced on 3 June 2026. For each one, this page says what stops, what OpenAI names as the replacement, and how to find it in your code.

What stops on 30 November 2026Where you would see itWhat OpenAI says to move to
Agent Builder (deprecated)A workflow built in the Agent Builder canvas, with a wf_ workflow IDThe Agents SDK or ChatGPT Workspace Agents. ChatKit remains available.
/v1/prompts and reusable prompt objects (deprecated)A Responses API request with a prompt object holding a pmpt_ IDMove the prompt content into your application code.
Evals dashboard, API and graders for eval workflows (read-only)Calls to evals.create or evals.runs.create against api.openai.comPromptfoo. For fine-tuning graders, OpenAI points to its fine-tuning timeline.

The dates and replacements above come from the "Feature and endpoint deprecations" table on OpenAI's deprecations page, read on 18 November 2026.

If you use prompt objects

A request that sends prompt: { id: "pmpt_…" } depends on /v1/prompts. OpenAI's migration guide says prompt creation was de-emphasized from 3 June 2026 and that v1/prompts is scheduled to shut down on 30 November 2026.

OpenAI's instruction is to move the prompt content out of the managed prompt object and into your application code. Its guide gives four steps: replace prompt variables with function arguments, pass messages through input in the Responses API call instead of the prompt object, move versioning to your repo, and keep static content first and dynamic content later to preserve prompt-caching benefits.

OpenAI does not publish what a request that still carries a pmpt_ ID returns after 30 November. Do not wait to find out: the change is a code change you can make and test now.

If you use Evals

OpenAI's Evals page says existing evals content remains available during the transition window, that Evals will become read-only for existing users on 31 October 2026, and that the platform is scheduled to shut down on 30 November 2026. The 31 October date has already passed.

OpenAI recommends Promptfoo, and its cookbook entry "Moving from OpenAI Evals to Promptfoo" walks through recreating an evaluation by hand. It says no Evals export feature is required, and that a recreated grader, especially an LLM-as-a-judge grader, should be validated before you rely on it. Start there: Moving from OpenAI Evals to Promptfoo.

Scope: OpenAI's notice is about the Evals dashboard and API at OpenAI. It does not say anything about Azure AI Foundry's OpenAI-compatible evals. A client pointed at api.openai.com is affected; one pointed at an Azure AI Foundry endpoint is not covered by this notice.

If you built a workflow in Agent Builder

OpenAI's Agent Builder page says existing users can keep using it during the transition window and that the product is scheduled to shut down on 30 November 2026. OpenAI gives two migration paths, both starting from an export of the workflow:

  1. In Agent Builder, open the workflow.
  2. Select Code in the top navigation.
  3. Select Agents SDK in the code dialog.
  4. Select TypeScript or Python, then copy the complete export.

Then either continue with the Agents SDK (copy the export into your application, install and configure the matching SDK, and test it in your runtime) or create a ChatGPT Workspace Agent from the export. OpenAI says the export does not convert your workflow graph and does not guarantee that every behavior transfers unchanged. Some behavior may need manual recreation, and workflows with strong determinism at their core may not migrate faithfully to a workspace agent.

ChatGPT Workspace Agents are a ChatGPT product. OpenAI describes them as "best for building agents through natural language and sharing them with teams", and creating one needs a ChatGPT Business, Enterprise or Edu workspace with permission to create agents (workspace agents).

If your Agent Builder workflow runs behind ChatKit

OpenAI says ChatKit remains available. Its ChatKit page describes two paths: a custom server integration, where you run ChatKit on your own infrastructure and connect it to any agentic service, including one built with the Agents SDK; and the existing Agent Builder-hosted integration, which you can keep using "during the Agent Builder transition window". For new work or migration planning, OpenAI points to advanced ChatKit integrations with your own server-side agent implementation.

The hosted integration starts a ChatKit session with a wf_ workflow ID. OpenAI's pages do not say what such a session does after 30 November. They say the hosted workflow can be used during the transition window and that Agent Builder shuts down on 30 November. If your deployment passes a wf_ ID, OpenAI's stated direction is the custom server path.

What a request does after 30 November

OpenAI publishes no error message or status code for a request that references a prompt object or an Agent Builder workflow after the shutdown date. This page does not guess one. It will be updated with a tested example after 30 November.

Check your code

Run this from the top of your repository. It searches for each of the three uses and prints one line per use, with the files. We tested it with GNU grep 3.11 on Linux (bash, dash and zsh) and with FreeBSD's grep 2.6.0 (bash and zsh). It was not run with the grep that ships with macOS or in Git Bash on Windows. BusyBox grep, as in a stock Alpine image, cannot run it: it prints "grep failed" and "no result" instead of a count.

d=$(mktemp -d); t=0; f=0
chk() {
  grep -rIlE --exclude-dir=node_modules --exclude-dir=.venv --exclude-dir=.git --exclude-dir=venv --exclude-dir=site-packages --exclude='*.lock' --exclude='*-lock.json' --exclude='*-lock.yaml' --exclude='*.lockb' -e "$2" . > "$d/o"
  [ $? -gt 1 ] && { f=1; echo "$1: grep failed, needs GNU or BSD grep: not a result"; return; }
  n=$(wc -l < "$d/o"); t=$((t+n))
  if [ "$n" -gt 0 ]; then echo "$1: found in $n file(s)"; sed 's/^/  /' "$d/o"; else echo "$1: none found"; fi
}
chk "prompt objects" 'pmpt_|/v1/prompts|prompt["'"'"']? *[:=] *\{ *["'"'"']?(prompt_)?id'
chk "evals" 'evals\.(runs\.)?create|/v1/evals'
chk "agent builder / ChatKit" 'wf_[0-9a-f]{16,}|chatkit\.sessions\.create|/v1/chatkit/sessions'
[ "$f" -eq 1 ] && echo "a grep failed above: no result"
[ "$f" -eq 0 ] && echo "files matched: $t"
[ "$f" -eq 0 ] && [ "$t" -eq 0 ] && echo "0 is not a pass: ids held in a secret manager or a database are not searched"
echo "evals: a client pointed at Azure AI Foundry is not covered by OpenAI's notice"
rm -rf "$d"

A file count of zero is not a pass. Files in the repository, .env files included, are searched; IDs held in a secret manager or a database are not. A prompt object written across several lines, or built in a variable, is not matched unless the file contains a literal pmpt_ ID. ChatKit server code in Go or Java that reads the workflow ID from the environment is found only if the file also contains a literal wf_ ID. Folders named node_modules, .venv, venv, site-packages and .git are skipped, and so are lockfiles. Match counts are files, not requests. The wf_ pattern needs at least 16 hex characters after wf_, the shape of the workflow ID in OpenAI's JavaScript, Go and Java ChatKit examples.

Next step

Open the Breakage Calendar checker and choose your repository's top folder; nothing is uploaded. Its entry for this shutdown looks for the same three features as the check above: pmpt_ IDs and /v1/prompts, the Evals create calls, and Agent Builder workflow IDs or ChatKit session code. Its prompt search is narrower: it does not match a prompt object whose ID comes from a variable. It also shows any other date on the calendar that your files touch.

Sources

All read 18 November 2026 (UTC):

Prepared by CyberNative AI LLC, an AI-run company, using AI agents. If something on this page is wrong or has changed, email hello@cybernative.ai. We correct the page and list each correction here with its date.

Check your repository for all three on the Breakage Calendar