OpenAI Evals read-only October 31, 2026; Evals shutdown and Claude Sonnet 4.5 retirement November 30, 2026
Two November 30 Deadlines for AI Governance: OpenAI Evals Goes Read-Only and Claude Sonnet 4.5 Retires
Two vendor clocks that AI governance teams track reach the same date this month. OpenAI announced on June 3, 2026 that it is shutting down its Evals platform: existing evals became read-only on October 31, 2026, and the dashboard and API shut down entirely on November 30, 2026. Separately, Anthropic notified developers on September 30, 2026 that Claude Sonnet 4.5 is scheduled to retire on November 30, 2026, under its commitment to at least 60 days of notice before retiring a publicly released model. OpenAI notice periods are longer on paper, at least 6 months for generally available models and at least 3 months for specialized variants, with shorter periods possible where safety or compliance concerns require faster action. The operational risk sits in the overlap, as an analysis of model switching costs published this week spells out: a team that kept its test cases only inside the OpenAI Evals platform must migrate its testing setup at the same time as it may be forced to revalidate a workflow on a successor model, and the two migrations compound. A cheaper or newer model is a configuration change only in appearance; the real cost is revalidation, checking that a workflow still meets its agreed standard on the new model, with evidence that survives the switch: examples of acceptable and unacceptable work, the criteria used to judge them, the workflow configuration, and the results the current setup produces on those examples. Teams that hold that evidence in systems they control can re-run it on any successor; teams that held it only inside a retiring vendor platform start from reconstruction, under a deadline the vendor set.
Executives
Vendor retirement dates are governance dates. November 30 now carries two of them at once: the evaluation platform some teams use to prove their AI workflows work, and a widely used model those workflows may run on. The leadership question is narrow and answerable this week: for each production AI workflow, can a named owner show the test evidence, stored in our own systems, that would let us revalidate on a successor model inside a notice period. If the answer is no, that gap, not the model price list, is the real switching cost.
CIO / CISO
Export anything still held in OpenAI Evals before the November 30 shutdown, while the read-only window remains open, and store test cases with the workflow configuration and expected outputs in version control you own. Inventory workflows running on Claude Sonnet 4.5 and schedule revalidation against its successor with time to spare before the same date. Under the NIST AI RMF Measure and Manage functions, treat each forced migration as a recorded decision, owner, standard, evidence, and sign-off renewed, rather than a silent endpoint change.
Engineers
Run the evidence test now on one consequential workflow: gather five or more real past inputs with outputs you consider acceptable, plus any output you consider wrong, and record the results the current model produces on them today. That set becomes the regression pack for the successor model. When you switch, add a case for every new failure you discover, and keep inspecting production results for a period after deployment, because passing the existing cases does not prove the new model fails in no new ways.
Everyone (non-IT)
If an AI tool you use at work behaves slightly differently in December, a scheduled model retirement like this one is a common reason: vendors replace models on published dates, and responsible teams re-check their tools afterward. You do not need to do anything, but reporting a clear change in output quality to your IT or AI governance contact genuinely helps, because drift is easier to catch from real examples than from dashboards.