OpenAI Evals migration: what to check before November 30
OpenAI’s published schedule puts existing Evals into read-only mode on October 31, 2026, ahead of a November 30 shutdown of the Evals dashboard and API. This is a planned transition announced on June 3, not a new product announcement.
The official deprecation schedule also lists November 30 for reusable prompt objects and Agent Builder. ChatKit remains available. ByteForward has not tested account-level enforcement of the read-only transition.
Three dependencies to check before November 30
| Dependency | Published deadline | Migration path |
|---|---|---|
| Evals dashboard and API | Read-only October 31; shutdown November 30 | Recreate tests in Promptfoo and check scoring behavior. |
| Reusable prompt objects | November 30 | Move managed prompt content into application code. |
| Agent Builder | November 30 | Export Agents SDK code; use the SDK or recreate a workspace agent. |
Evals migration is a rebuild, not a history transfer
OpenAI’s Promptfoo migration guide describes manually rebuilding an evaluation from its prompts, model provider, test cases and scoring rules. It does not require an Evals export feature. Configuration lives alongside code and can run locally or in continuous integration.
A fresh Promptfoo run is separate from past OpenAI Evals runs. Recreated graders need validation, particularly similarity scoring and model-based judges, because matching inputs do not guarantee identical scores. Custom tools or agent workflows may require their own configuration.
For a practical first pass, choose one evaluation whose expected behavior your team already understands. Include examples that should pass and examples that should fail. Compare individual decisions before adopting a new aggregate score as a release gate. A dashboard that opens successfully is not evidence that the replacement catches the same regressions.
Prompt references need their own review
The prompt-object guide moves managed prompt content into application code. Review applications that submit a stored prompt ID, version and variables rather than assuming an Evals migration also fixes that dependency.
Keep the intended prompt version and representative variable values together during that review. Otherwise, a migration can quietly change both the evaluation system and the instructions being evaluated, making a score difference harder to diagnose.
Agent Builder exports still need runtime testing
For Agent Builder, OpenAI’s migration guide starts with the workflow’s Code dialog: choose Agents SDK, then copy the complete TypeScript or Python export. Teams can continue in their own application or use the export to help recreate a ChatGPT Workspace Agent.
The workspace route requires an eligible Business, Enterprise or Edu workspace, agent access and creation permission. OpenAI warns that the process does not automatically convert the graph or guarantee unchanged behavior. Review tools, triggers, control flow, authentication and permissions separately, especially for workflows that depend on deterministic execution.
The useful milestone is a validated replacement, not simply copied code. Assign an owner to each dependency, record which representative cases passed and keep a clear distinction between a completed migration and a scheduled one.
Editorial illustration by ByteForward, generated with AI. OpenAI wordmark: OpenAI. Conceptual imagery, not a product interface or an export of historical runs.
