OpenAI will require AI gateways and routers to get its approval
OpenAI's Nov 20 rule makes it the gatekeeper for any gateway reselling its models — plus SafetyIDs, Mistral's Voxtral 3, and an open bio-agent atlas.
Copy markdown
The rule: a gatekeeper for every gateway
From Nov 20, any router, gateway, or reseller offering direct access to OpenAI models must get OpenAI's review and approval first. If you run or lean on an OpenAI-backed gateway, that access is now conditional, not automatic.
One restriction now follows you across providers
Partners and resellers must mirror OpenAI's safeguards, and a block from one provider is meant to carry across every service offering access. Routing around a restriction by hopping gateways is the behavior this is built to end.
SafetyIDs: a signal your stack may soon forward
OpenAI is building SafetyIDs — anonymized per-user identifiers customers can supply so it can tell end users apart during investigations without learning who they are. No verification gate today, but plan for gateways to pass end-user signals upstream.
Also today: Mistral's Voxtral 3 voice stack
Mistral shipped Voxtral 3 — Transcribe (STT), TTS (~$13 per 1M chars), and a Voice Agent SDK — on its EU-hosted API or self-hosted on your own infra. A deployable, European alternative for building real-time voice agents.
Also today: an open 'reasoning atlas' for agent fan-out
OpenAI, PerturbAI and Scripps ran thousands of GPT-Rosalind agents over a whole-brain CRISPR atlas in 32 minutes, publishing the reasoning atlas (GitHub) and dataset (Hugging Face). A working template for fanning agents across one huge dataset.