OpenAI Evals Becomes Read-Only October 31: The Agent Builder and Prompt Shutdown Checklist

30 October 2026 | Original announcement: 3 June 2026 | Status: Scheduled, not yet confirmed effective. If your AI team uses OpenAI's hosted Evals dashboard, tomorrow matters. Existing evaluations are scheduled to become read-only on 31 October 2026, followed by shutdown of the Evals dashboard and API on 30 November 2026. Agent Builder and reusable prompt objects also have 30 November retirement dates. These changes were originally announced in June, not today. The immediate news is the approaching operational cutoff. Check the official deprecations register before planning your cutover.

Three deadlines, three distinct impacts

31 October: Evals read-only. Teams relying on hosted evaluations should preserve the configurations, cases, grader definitions and results they need for migration while the existing interface remains writable. OpenAI confirms the read-only date in its Evals developer guide. Read-only is not the same as shutdown: do not tell colleagues that the entire service disappears tomorrow.

30 November: Evals dashboard and API shut down. Scheduled evaluation jobs, regression gates and integrations that depend on the legacy hosted Evals API need an alternative before that date. OpenAI's Promptfoo migration walkthrough shows how to recreate prompts, providers, test cases and assertions manually. It does not promise an automatic migration or an export tool.

30 November: Agent Builder and reusable prompt objects retire. The visual Agent Builder product is scheduled to shut down, but ChatKit remains available. A ChatKit frontend backed by an Agent Builder-hosted workflow still needs a tested replacement backend. Separately, the v1/prompts API and reusable managed prompt objects are scheduled to shut down. These are specific product retirements, not a retirement of the Responses API or ChatGPT. See the ChatKit documentation and prompt migration guidance.

What enterprise teams should do today

Start with a dependency inventory rather than forwarding a generic AI warning. Find CI jobs that call the Evals API, applications that depend on managed prompt IDs, and ChatKit deployments configured with Agent Builder workflow IDs. For each one, identify an owner, customer-facing impact, replacement architecture, verification evidence and migration deadline. A tool being deprecated does not prove your organization uses it, and the right response is to establish exposure before allocating work.

For Evals, preserve the evaluation definitions and the information needed to reproduce them, observing your organization's data-handling requirements. OpenAI's official migration example explains how to create a Promptfoo configuration, execute a new evaluation and compare results. This is a manual recreation. Similarity metrics and model-graded judgments may differ, so a migrated test is not validated merely because its configuration file loads.

For Agent Builder, the migration guide describes exporting a workflow as Agents SDK code or recreating its intent with ChatGPT Workspace Agents. Neither path guarantees identical behavior. Workspace Agents require eligible workspace access; they should not be described as a universal API backend replacement. Revalidate tools, triggers, authentication, approvals, error handling and output contracts in a non-production environment before any traffic switch.

For reusable prompts, OpenAI recommends moving prompt content into application code, turning dynamic variables into explicit inputs, and maintaining versions through the usual software development process. Its prompt-object migration guide provides examples. Review your application's prompt caching and deployment tests after the change rather than assuming identical results from copied text.

A practical five-point cutover checklist

1. Discover: Search configuration, repositories and CI pipelines for hosted Evals calls, Agent Builder workflow references and reusable prompt IDs. Validate findings with system owners.
2. Preserve: Retain authorized copies of evaluation test cases, grader definitions, prompts, configurations and baseline results while accessible.
3. Recreate: Build equivalent tests in the selected replacement and compare normal, edge and failure cases. Investigate scoring differences.
4. Verify security: Retest permissions, service credentials, tool approvals, private-data handling and end-to-end ChatKit behavior after migration.
5. Close the dependency: Assign a migration owner and cutover date, record the fallback plan and confirm that production monitoring no longer depends on a retiring service.

What this announcement does not mean

OpenAI is not shutting down ChatGPT, ChatKit, the Responses API or the practice of AI evaluation. The dates concern named hosted developer products. The original deprecation notice was published on 3 June 2026; this article highlights an approaching effective date, not a fresh 30 October policy announcement. As of publication, the 31 October read-only milestone is scheduled, not independently confirmed as already executed. Consult the live official schedule for changes.

Bottom line: Teams that still use legacy OpenAI Evals should treat 30 October as the last preparation day before the scheduled read-only transition. Then use November to complete and test the Agent Builder, Evals and prompt-object migrations. The real risk is not missing a product announcement; it is losing a dependable evaluation or workflow control in a production AI system.