Originally published on Litmiq. Updated 13 November 2026. Sources on that page were read the same day.
OpenAI Evals became read-only on 31 October 2026 and is scheduled to shut down on 30 November 2026 (dashboard, API, and graders for eval workflows). OpenAI recommends migrating to Promptfoo. Dates and wording should be checked against OpenAI's deprecations page, which is authoritative if anything here disagrees.
This is a short export-first path. The full guide, including the export script and an in-browser converter, is at litmiq.com/migrate-from-openai-evals.
The short answer
- Export first. There is no export button, but the Evals API still returns your evals, runs, and output items while the product is read-only. Save all of them, read-only, before 30 November 2026.
-
Convert. The Litmiq guide includes a free, in-browser converter that turns that export into a
promptfooconfig.yaml, a CSV, or a Litmiq suite. - Re-check scores. Similarity and model-graded scores will not match OpenAI's exactly. Re-run a sample and re-tune thresholds before you trust them.
A close, separate deadline from the same deprecations page (read 13 November 2026): gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro are listed as shutting down on 11 December 2026. If an eval pinned one of those, re-baseline on the replacement OpenAI names.
Timeline (as published by OpenAI, read 13 Nov 2026)
| Date | What happens |
|---|---|
| 3 June 2026 | OpenAI publishes the deprecation notice for the Evals dashboard, API and graders for eval workflows, and for /v1/prompts reusable prompt objects. |
| 31 October 2026 | Evals become read-only for existing users. |
| 30 November 2026 | The Evals platform is scheduled to shut down. OpenAI recommends Promptfoo. Reusable prompt objects are deprecated the same day — move prompt text into your app code. |
| 11 December 2026 | Separate but close: several GPT-5 / o3 models shut down. Re-baseline if you pinned them. |
If OpenAI changes these dates, their pages win.
Where to go after you export
- Promptfoo — this is OpenAI's recommended replacement, and the right default for CLI and CI workflows.
- A browser UI — if you want to inspect cases side by side without installing anything, you can load a converted suite in Litmiq. Litmiq is free, you bring your own OpenAI key, and it does not replace Promptfoo as a merge gate. I built Litmiq; the longer guide is explicit about those limits.
DeepEval and other code-first tools remain reasonable if that is already your stack. The point is not to switch tools twice: export once, then pick the workflow you will actually run.
What will not come across cleanly
Model-graded and similarity scores from OpenAI's graders are not something you should copy-paste as pass/fail thresholds in a new tool and trust. Re-run a sample. Treat old scores as history, not as a grade book that still applies.
The in-browser converter and the export steps (including a worked IT-ticket example) are on the canonical guide. I am not pasting a GitHub repo here, because the public converter repository is not published yet; use the guide on the site.
A short checklist
- Export every eval, run, and output item you still care about, while the API is read-only and still answering.
- Keep the raw export, even after you convert it.
- Convert to Promptfoo YAML and/or a CSV / Litmiq suite.
- Re-run a sample on the models you will actually use after 11 December 2026.
- Re-tune thresholds. Do not ship a CI gate on un-checked migrated scores.
- Move any reusable prompt objects out of
/v1/promptsand into your own source.
If you find an error in the dates or the export steps, the canonical page is the best place to flag it — that is the document I will correct first.
Top comments (0)