Tools2 min read

OpenAI Evals enters its read-only stage: check business testing continuity

OpenAI's schedule puts existing evals in read-only mode from October 31. Check how your team will test AI changes before the platform closes.

Prepared by the NADALI editorial team using AI

AI illustration: a hand fits a turquoise panel onto a server beside a conceptual business website window.
Nadali — original AI-generated editorial illustration created with OpenAI image generation; reused from Nadali archive. · Original AI-generated illustration created for Nadali; use subject to applicable OpenAI terms; no exclusive copyright claim; no third-party input images used. · 1440 × 960

Author: Nadali. Prepared using AI.

Under OpenAI’s current schedule, existing Evals evaluations become read-only on October 31, 2026. The Evals dashboard and API are scheduled to close on November 30. Deprecation was announced on June 3: the new stage now concerns existing evaluations, rather than a newly announced shutdown.

Who needs this: SMB owners whose developer or contractor uses OpenAI Evals specifically to check a business assistant’s responses. This is not an announcement ending every way to evaluate AI.

Questions for your technical team

Our editorial recommendation: ask where the next test will run after a model, instruction or connected-tool change. If the answer relies on the old Evals dashboard, agree on a migration owner and acceptance criteria.

Preserve available input examples, expected results, scoring rules and recent reports in an approved repository. Do not automatically transfer customer data to another service. Record missing components separately: a list of test names alone cannot recreate an evaluation.

Migration needs its own checks

OpenAI’s guide describes manually recreating an evaluation in Promptfoo, not a guaranteed complete export. Prompts, models, test cases and criteria need mapping. A new run is separate from previous runs; numerical scores may differ, and recreated graders need validation.

A practical sequence for the business commissioning the work:

  1. Choose one important process and approved test data.
  2. Ask the team to explain each discrepancy between the old and recreated evaluation.
  3. Separately check cases where an error must block release.
  4. Assign someone to accept the result before the next change to the working assistant.

This is an editorial action sequence, not a Nadali test result or a recommendation to buy another product.

Limits and takeaway

We checked the public schedule and documentation, not the state of an individual account. This article does not promise complete history transfer or identical scores. Costs and access for a replacement environment need separate approval.

The takeaway: check testing continuity now so that a supplier’s tooling change does not leave your business assistant without its control checks.

Sources read November 1, 2026 UTC. Cover: a reused conceptual AI illustration of technical maintenance, not the Evals interface. 

Sources and verification

  1. OpenAI API deprecations — Evals platform ↗
    Source checked: 2026-11-01
  2. Moving from OpenAI Evals to Promptfoo ↗
    Source checked: 2026-11-01