OpenAI Evals Is Shutting Down: Your Options Before 30 November
Last checked 8 November 2026.
OpenAI's Evals platform shuts on 30 November 2026. If you kept test cases and graders there, you have until then to get them out and decide where they go next. This guide covers what's closing, what to export, and three ways to keep testing.
Want to keep testing in a browser? Bring your Evals or Promptfoo files into a Prompt test, or paste a prompt and we'll draft the cases. Test your prompt →
What's closing, and when
| Date | What happens |
|---|---|
| 3 June 2026 | OpenAI announced the deprecation of the Evals platform. |
| 31 October 2026 | Existing evals became read-only. |
| 30 November 2026 | The Evals dashboard and API shut. Reusable prompts (v1/prompts) and Agent Builder shut on the same day. |
OpenAI's migration guide is “Moving from OpenAI Evals to Promptfoo”. Source: OpenAI's deprecations page, checked 8 November 2026.
Anthropic made a similar change earlier in the year. Legacy Workbench access ended on 17 August 2026. Saved prompts, variables and evals aren't supported in the updated Workbench and Playground, and the experimental generate_prompt, improve_prompt and templatize_prompt APIs were retired the same day. Source: Claude release notes, entry dated 17 July 2026.
1. Export your data first
Whatever you choose next, save what you have before 30 November. Read-only means you can still read it.
- Your test rows. If you built an eval from an uploaded JSONL file, keep a copy of that file. If the rows were inline, read the eval run back as JSON: its data source holds the rows and the message template.
- Your prompts. Copy the full text of each prompt and message template, including the system message. If you used reusable prompts, copy those out too: they close the same day.
- Your graders. Write down each testing criterion: what it compared, and with what. A string check on a fixed word is easy to carry over. A model grader's instructions are worth keeping word for word.
- A baseline. Save the last results you trusted, so you can tell whether the new setup agrees with the old one.
We haven't checked every export screen, so follow OpenAI's own guide for the exact steps.
2. Choose where to test next
There's no single right answer. It depends on who runs the tests and where.
- Promptfoo is the route OpenAI's guide describes. It's a command-line tool: you write a config file listing prompts, providers and tests with assertions, and run it from a terminal or in CI. It suits developers who want tests in the repo next to the code.
- Your own script is fine for a handful of cases. It gets harder once you need several models, judged checks and a record of what changed.
- A browser tool suits people who want to test without setting up a project. Spring Prompt's Prompt test is one: paste a prompt, and it drafts test cases and checks for you, runs the models you pick side by side, and shows every answer with its cost. It runs models from many developers, so you can compare the replacement for your current model with others.
You can also use more than one. A config in CI catches regressions on every change; a browser test is quicker for choosing a model or rewriting a prompt.
3. Bring your evals into a Prompt test
In a new Prompt test, choose Import from my app, then paste or upload the file. These formats load straight in, with no model call:
| You have | What comes in |
|---|---|
An Evals JSONL file with input messages and an ideal answer | The system message, the user message as a template (wording every row shares stays fixed), one case per row, and the ideal answer as each case's expected answer. |
Evals dataset rows ({"item": {...}}) | The fields as variables, and answer fields (such as answer or ideal) as each case's expected answer. |
Chat transcripts ({"messages": [...]}) | The conversation as the input, and the last assistant reply as the expected answer. |
| An eval run as JSON, with its message template and rows | The template, with {{ item.field }} turned into a variable, the rows as cases, and string checks on fixed text as checks. |
A Promptfoo config (promptfooconfig.yaml or JSON) | Prompts, the first provider's model, tests as cases, and assertions such as contains, icontains, regex, is-json and llm-rubric as checks. |
A Prompt test takes up to 12 cases and 10 checks, so a large file is cut to the first ones, and the import tells you. It also lists anything it couldn't carry over: graders written as code, model graders, files referenced with file://, and images. You can ask it to draft checks to replace them. API keys in the file are removed before anything is stored, and provider settings aren't copied.
4. Check your models' dates while you're there
If you're moving your tests, it's a good moment to check the models under them. Claude Sonnet 4.5 retires on 30 November too, the gpt-5 family on 11 December, and gpt-4o and gpt-4.1 on 14 January 2027. Our model retirements page lists each date with its official replacement and source, and what to re-test before you switch.
Moving your evals and your model at once makes it hard to tell which change moved a result. If you can, carry the tests over first, check they agree with your last baseline, then change the model.
Bring your cases, or start from a prompt. Run the old and new model on the same inputs and read every answer side by side. Test your prompt →
Quick answers
When does OpenAI Evals shut down?
OpenAI announced the deprecation on 3 June 2026. Existing evals became read-only on 31 October 2026, and the Evals dashboard and API shut on 30 November 2026.
What does OpenAI suggest instead?
OpenAI's migration guide is called "Moving from OpenAI Evals to Promptfoo". Promptfoo is a command-line tool that you configure with a file of prompts, providers and tests.
Is anything else shutting down on 30 November?
Yes. OpenAI's reusable prompts (v1/prompts) and Agent Builder shut on the same day, according to its deprecations page.
Can I bring my OpenAI Evals data into Spring Prompt?
Yes. Paste or upload an Evals JSONL dataset or an eval run with its rows, or a Promptfoo config. Prompts, test cases and the checks we can run come in without a model call. Graders written as code, and model graders, are listed as left out so you can replace them.
Does Anthropic's Console still have evals?
Not as before. Legacy Workbench access ended on 17 August 2026, and saved prompts, variables and evals aren't supported in the updated Workbench and Playground.
