OpenAI Shuts Prompts API, Evals and Agent Builder on November 30: What to Move Before the Deadline
On 30 November 2026 OpenAI shuts down three platform features at once: the /v1/prompts endpoint with its reusable prompt objects, the Evals dashboard and API with its graders, and Agent Builder. Any team that stored prompts on the platform, ran regression tests in Evals or built a no-code workflow in Agent Builder has 11 days to move that work into code or into another tool, and OpenAI documents no bulk export for prompts.
Key facts
- 3 June 2026: OpenAI announced the deprecation of
/v1/prompts, the Evals dashboard, API and graders, and Agent Builder, all with a shutdown date of 30 November 2026. - 31 October 2026: existing evals became read-only, according to eCorpIT's shutdown tracker; OpenAI's deprecations page lists Evals with the status "Read-only".
- Replacements named by OpenAI: prompt text in your application code, Promptfoo for evals, and the Agents SDK or ChatGPT Workspace Agents for Agent Builder. ChatKit remains available.
- 18 November 2026: a new notice retires GPT-5.2, GPT-5.4 and GPT-5.4 mini with EU data residency on the fast or priority service tier, effective 11 January 2027.
- 11 December 2026:
gpt-5,gpt-5-mini,gpt-5-nano,gpt-5-pro,o3ando3-proshut down.
Reusable prompts: copy the text out yourself
Prompt creation was de-emphasized from 3 June, and the endpoint is scheduled to shut down on 30 November. OpenAI's migration guide describes no export method. It tells you to move the prompt content into your codebase: a small prompts module with one builder function per prompt, prompt variables turned into typed function arguments, and the resulting messages passed as input to the Responses API instead of a prompt object with a pmpt_ ID.
Two details from the guide matter for cost and control. Put static content first and dynamic content last, because prompt caching only hits on an exact prefix match. And version prompts in git with pull request review and tests, which replaces the version history the platform used to keep.
Evals: rebuild in Promptfoo, keep your own records
The Evals dashboard, API and graders for eval workflows shut down on 30 November, and OpenAI recommends Promptfoo. The cookbook covers recreating an eval by hand in a promptfooconfig.yaml file, where OpenAI's testing criteria become Promptfoo assertions and metrics. It also warns that a recreated grader, an LLM-as-judge grader in particular, needs validation before you trust it for regression decisions.
An OpenAI support reply on the developer forum describes two export types from the eval menu: a runnable Promptfoo config with prompts, models, test cases and assertions, and a JSONL file that only holds the results of a completed run. Not every eval type supports the Promptfoo config export. OpenAI has not published what happens to eval run history after shutdown, so treat the 30th as the last day you can read it.
Graders used in fine-tuning follow a separate timeline: active self-serve fine-tuning customers can no longer create new jobs from 6 January 2027, while inference on fine-tuned models continues until the base model is deprecated.
Agent Builder: export code, then test what changed
Agent Builder can export a workflow as Agents SDK code in TypeScript or Python from the Code menu. OpenAI's migration guide says this "does not convert your workflow graph or guarantee that every behavior transfers unchanged". The other route, a ChatGPT Workspace Agent, needs a Business, Enterprise or Edu workspace, and the guide warns that workflows with strong determinism at their core may not migrate faithfully.
The open question is live traffic. OpenAI has not said whether already-published workflows keep running after the 30th, and its ChatKit guide now steers teams away from the hosted path that uses an Agent Builder workflow ID. If a customer-facing chat widget depends on a hosted workflow, plan as if it stops on 30 November.
The EU fast tier notice and the December model retirements
The newest entry on the deprecations page, dated 18 November, affects EU teams paying for speed: GPT-5.2, GPT-5.4 and GPT-5.4 mini with EU data residency and service_tier: fast or priority end on 11 January 2027. The options are the same models on Standard processing (service_tier: default) with EU residency, which keeps residency but drops the speed tier, or a move to the GPT-6.2 family. Before that, on 11 December, the GPT-5 base family and o3 models retire in favour of gpt-6.1-sol, gpt-6.1-terra and gpt-6-luna.
What to do this week
- Search your code for
pmpt_IDs and thepromptparameter, copy every live prompt version into the repository and switch calls toinput. - Export each eval as a Promptfoo config where the menu allows it, download JSONL results for the rest, and save the scores you report on. Our guide to building a small eval covers what a minimal replacement suite needs.
- Export Agent Builder workflows to Agents SDK code, run them against the same inputs, and compare outputs before the 30th.
- List every OpenAI model ID and service tier in production against the 11 December and 11 January dates.
- Check other vendors for the same date: Anthropic also removes fast mode on Claude Opus 4.8 on 30 November.
More coverage of platform changes is in the article archive.
Questions people ask
Can I export my OpenAI evals before they shut down?
Partly. OpenAI support describes a Promptfoo config export and a JSONL export of completed results, but not every eval type supports the config export, and the cookbook assumes you recreate evals by hand.
What replaces OpenAI reusable prompts?
Your own code. OpenAI's guide moves prompt text into builder functions in your repository, with variables as typed arguments and versions managed in git.
Will my ChatKit embed keep working after 30 November?
ChatKit itself remains available. OpenAI has not confirmed whether embeds backed by a hosted Agent Builder workflow keep working, so move them to a server-side agent first.
What else is OpenAI retiring soon?
The GPT-5 family and o3 models on 11 December 2026, EU fast and priority processing for GPT-5.2 and GPT-5.4 on 11 January 2027, and new self-serve fine-tuning jobs for active customers on 6 January 2027.
Sources