OpenAI requires approval for gateways
On 21 October 2026 OpenAI announced that its updated Usage Policies, effective 20 November 2026, will require services that give direct access to its models, such as model routers and AI gateways, to get OpenAI's review and approval before offering that access. The post also introduces SafetyIDs, identifiers that customers can pass so that OpenAI can tell individual end users apart in an investigation without learning who they are.
Why it matters
The post moves part of the control to the intermediaries between OpenAI's models and their users: services that offer direct access, such as routers and gateways, must be reviewed and approved, and partners must keep safeguards consistent with OpenAI's, investigate violations and coordinate on enforcement. SafetyIDs are meant to let OpenAI act against one end user instead of disrupting a whole customer. The page gives no list of approved services, no review criteria and no date for SafetyIDs beyond "working with customers to introduce" them.
What the post says. "Over the past few months, we've worked with customers and partners to align safety policies across services that provide our models." The policies apply whether people reach the models directly through OpenAI, through partners that host them or through resellers; partners must maintain safeguards consistent with OpenAI's, investigate potential violations and take action. The updated Usage Policies, effective 20 November 2026, "clarify the rules for safe model development and sensitive uses of AI" and add the approval requirement. Private Safety Processing is named as an example of safeguards that leave customers control of sensitive data. Misuse can come through end users, compromised accounts or leaked API keys. SafetyIDs are identifiers customers can provide "to help us distinguish between individual end users in an investigation without OpenAI learning those users' identities". OpenAI says enforcement rests on documented conduct and risk and that restrictions are revisited when conduct, commitments or safeguards change. A parallel. Anthropic's platform release notes for 20 October, a day earlier, describe user profiles through which a safety enforcement action can apply to one end customer instead of a whole organization. It is a similar mechanism; neither page links it to the other. What the record does not claim. The text of the updated Usage Policies was not read; the page has no approval procedure, criteria or list of approved intermediaries. The date of 20 November is the post's word; it has not yet arrived.