OpenAI's August–October 2026 model retirements affect applications that still route requests to removed identifiers. This is separate from migrating the Assistants API: a supported API endpoint can still fail when the requested model has been retired. The practical task is to find the actual model selections throughout an application, including fallback routes and background jobs.
Chat alias removal and cyber-model shutdown have different dates
The official deprecation schedule records August 10 removal of gpt-5.2-chat-latest and gpt-5.3-chat-latest, with GPT-5.6 Sol as the replacement. It also records September 11 deprecation of gpt-5.4-cyber and October 1 removal; its replacement depends on the cyber-model access actually available to the organization.
A replacement recommendation is not a promise of identical behavior. It also does not grant access to a restricted model. Check the current model and account requirements, then evaluate a supported candidate against the tasks the retired model handled. Preserve the distinction between an access problem, an unsupported identifier, and an application integration failure.
Find model selection outside the main request path
Search the application's configuration and direct callers, then inspect relevant routing tables, scheduled workers, and retry paths. A chat screen can use a current model while a nightly document processor or a failure fallback still requests a retired alias. Record the owner and intended purpose of each selection before replacing it.
Logs should identify the requested model and failure category without exposing user content or credentials. Group errors by the caller rather than assuming that one visible failure represents every affected workflow. This makes the recovery plan smaller and more reliable than changing all model names without understanding their role.
Compare a maintained replacement on the actual workload
OpenAI's model-selection guidance recommends comparing the same inputs and keeping a setting that meets the required quality level. Use that as a starting point, then define application-specific acceptance conditions: valid output shape, correct tool arguments, grounded answers, completion time, and total cost.
Include difficult cases that the old integration depended on. A workflow with a strict parser needs more than a fluent response; an agent needs to handle failed tools and changed instructions without performing an unintended action. Keep the prompt and surrounding tools stable during the first comparison so the result is attributable to the model change.
A fallback must itself remain available
Test failure handling explicitly. A route that retries a removed model under another retired alias cannot recover service. Select only candidates that the account can use, and make fallback conditions visible. Bound retries and preserve the application's existing permissions and checks.
After the cutoff, rollback means returning to a verified supported configuration, rather than restoring the unavailable identifier. Retain the previous integration settings for comparison, but distinguish a historical record from a configuration that can still run.
Close the migration with a workflow result
Verify each affected feature from request through rendered output and any saved state. Check the quiet background paths as well as the main interface, then monitor errors and correction effort during a limited rollout. A completed migration has no active caller selecting the retired identifier and preserves the task's acceptance conditions. A successful provider response alone cannot establish that the user received the right result.