On 21 October 2026, OpenAI added project and organization controls for permitted API service tiers. That makes a fallback policy worth checking: switching away from Flex must respect both the allowed tiers and the cost you agreed to. This is a practical review of that dated change, not a new Flex launch.
Flex processing trades lower prices for slower responses and possible resource unavailability. It can suit a draft report needed tomorrow more than an answer someone is waiting for now. Check that the selected model supports it; do not infer support from the model’s ordinary API availability.
Define the fallback before submitting
OpenAI documents a 429 Resource Unavailable response for Flex capacity shortages. That particular outcome is not charged. It is not a reason to assume that every 429 has the same cause or billing treatment.
For a fictional overnight report, choose an end time and a finite retry policy. If Flex remains unavailable, decide whether to leave the report pending or retry with standard processing. Removing the Flex setting or using the documented default route can change cost; do not hide that decision inside an automatic retry.
Inspect the finished job
Check the actual processing mode and usage returned by the service, then verify the report’s completeness. A successful fallback proves neither the original price nor the truth of the content. Keep the original error and final result linked so you can explain why the run cost what it did. No paid request is necessary to design this policy.
The current guide uses GPT-6 Astra in its Flex example; supported models and prices are listed on the current Flex pricing page. Check that list for your model and permitted region. A disallowed tier produces a documented 400 error, not the Flex capacity error. Setting service_tier to auto or omitting it uses the project’s resolved mode; it is not a guarantee of your intended standard price. Ask the administrator to confirm policy rather than changing restrictions automatically.