The 30-Second Brief
- Source report OpenAI’s published Evals retirement timeline sets October 31, 2026 as the read-only milestone and November 30 as the scheduled dashboard and API shutdown. [1]
- Analysis Teams should preserve their evaluation evidence and prove that replacement tests support the same release decisions.
What Changed
The October 31 milestone has now passed. This is an operational transition announced on June 3, rather than a newly announced retirement. OpenAI’s current documentation includes eval-workflow graders in the transition and points to Promptfoo as a migration path. Fine-tuning has a separate timeline. [1]
Why It Matters
Analysis Moving a test suite is more than copying prompts. A release decision may depend on the exact sample set, scoring rubric, model configuration and exception history. If those details change unnoticed, a new green result can appear comparable to an old green result while measuring something different.
Start with one representative evaluation that actually gates a deployment. Preserve the inputs, expected outcomes, grader instructions, thresholds and relevant historical results using supported access methods. Record which model versions and settings produced the baseline. Rebuild that test in the chosen replacement, then inspect disagreements at the individual-case level before relying on aggregate scores. A deliberate rubric improvement is useful; an undocumented scoring change is harder to audit.
For teams with many experiments, separate essential release gates from exploratory work. Give each required gate an owner and a migration acceptance criterion. Keep sensitive examples in approved storage with the same access restrictions as the original material. Completing an inventory early leaves time to investigate missing artifacts or unavailable export paths before the deadline.
Evidence & Uncertainty
This brief reports OpenAI’s documented schedule, checked November 2; BriefSpire has not tested an affected account or verified an export workflow. The source does not establish feature parity for a particular replacement. Priority: Important reflects a scheduled dependency change that can disrupt evaluation workflows, not a newly disclosed security incident.
Action Center
Selections remain in this browser.
Evidence Explorer
Is November 30 already a completed shutdown?
No. It is the scheduled shutdown date in the current documentation. The October 31 read-only milestone is the one that has passed. [1]
Does a migration recommendation prove equivalent scores?
Analysis: No. Teams should validate their own datasets, scoring rules and decision thresholds. A tool recommendation alone does not establish comparability.
Sources & Transparency
Checked . This is a primary vendor timeline, not independent testing of account behavior. Migration checks are BriefSpire analysis.
- OpenAI API: Deprecations — Evals platform — announcement dated June 3, 2026; milestones October 31 and November 30, 2026.