Model choice is now a data retention decision

An AI request can pass every technical check and still cross a policy line. The primary model may discard a sensitive prompt after processing while its fallback keeps the same prompt for 30 days. Both routes can return a good answer. Only one may satisfy the organization's data rules.

That distinction became concrete this summer. OpenAI announced on August 19 that eligible API customers can use frontier models without prompts or responses being retained after processing. Its design keeps customer content on infrastructure the customer controls while sending OpenAI narrowly defined safety signals. Images flagged as potential child sexual abuse material remain an exception for legally required review and reporting.

Anthropic's policy requires 30-day retention for Fable 5, Mythos 5, and future designated Covered Models, including organizations that previously used zero retention. The storage arrangement depends on the platform. Data retained through Amazon Bedrock or Google Cloud Agent Platform stays in that cloud environment. Other Anthropic models keep their existing terms.

Put the rule where requests are routed

Most model registries track capability, latency, region, and price. They now need the effective retention policy as well. For each deployment, record whether zero retention is available, where retained content lives, how long it stays, which exceptions apply, and when someone last verified the terms.

The router can then make an enforceable decision. A request containing an HR case file might be allowed to use two deployments and forbidden from three others. If an approved route is unavailable, the system should stop or ask for an authorized exception. Silent failover to a model with different terms breaks the data decision before the model generates its first token.

Find every other copy

Provider retention covers one part of the system. Conversation history may sit in the product database. Retrieved documents may be copied into a vector store. Prompts and tool results often appear in traces, support logs, or evaluation datasets. Those copies need their own purpose, owner, retention period, and deletion path.

The policy record also needs a date. Model terms can change by provider, account, platform, and model version. Engineering should keep the route enforceable and visible; legal and security teams still decide what the organization can accept.

One important detail is still moving. OpenAI's Private Safety Processing is a preview, with rollout and a technical white paper planned for September 2026. Axios reported the policy split, but the provider documentation remains the implementation authority. Any route that depends on the preview should stay provisional until the technical details arrive.