OpenAI disbanded the team meant to catch this exact problem
Issue #016 · 4 min read
OpenAI disbanded the team meant to catch this exact problem
Preparedness team gone, Stripe buys an AI gateway for $7B+, and Copilot adds two more models
The big story
OpenAI had a team whose entire job was to check whether its models could help someone build a bioweapon, run a large-scale cyberattack, or manipulate people at scale before the model shipped.
That team is gone. OpenAI has reportedly disbanded its preparedness team — The Verge.
The timing is the story. This lands the same week The Verge is writing that rogue AI agents "aren't science fiction anymore," and the same week Anthropic's CEO is out telling reporters the AI backlash is "fundamentally a crisis of trust" — TechCrunch.
One company is cutting the team that checks for worst-case behavior.
Another is doing press about how nobody trusts the industry to check for worst-case behavior.
For a PM, the lesson isn't abstract. If you build on top of a foundation model, the vendor's internal safety org is part of your risk surface whether you asked for it or not. Worth knowing who's still staffing theirs before you bet a roadmap on it.
What shipped
Stripe is reportedly buying OpenRouter, the AI gateway startup, for more than $7 billion. OpenRouter's own CEO has described it as "Stripe for AI" — TechCrunch.
If you route model calls through OpenRouter today to avoid locking into one vendor, that neutrality just got a new owner with its own incentives. Worth watching what happens to pricing and model access once the acquisition closes.
GitHub Copilot picked up two more models this week: Grok 4.6 and Gemini 3.7 Flash. Grok 4.6 is pitched at agentic coding and multi-step workflows — GitHub. Gemini 3.7 Flash is aimed at web and app development — GitHub.
Neither is a reason to switch anything today. But if you're already choosing models per-task inside Copilot, that's now a five-or-six-way decision instead of a two-way one, and someone on your team should own that choice explicitly rather than let it default.
OpenAI shipped Ultrafast mode for GPT-5.6 Sol, running up to 14 times faster and hitting 750 output tokens per second, powered by Cerebras — OpenAI. If your product has a latency complaint in the backlog that you've been blaming on the model, this is worth a quick benchmark before you assume it's unfixable.
What I'd actually do this week
- If you're a Vercel AI Gateway or OpenRouter customer, ask your infra lead what changes for you if the Stripe deal closes — pricing, uptime guarantees, or nothing. Get the answer in writing now, not after a surprise invoice.
- Pull up whichever foundation model vendor you depend on most and check publicly whether their safety/preparedness function still exists in the form it did six months ago. If you can't find out, that's itself worth flagging to your team.
- If you're running production traffic through GPT-5.6, test Ultrafast mode against your current setup on one real workflow. 14x is a big enough claim to verify cheaply before you plan around it.
Hit reply if you've already moved traffic through Ultrafast or you're rethinking your gateway setup — want to hear what you're seeing.
Tools mentioned
- OpenRouter Stripe acquisition
- Grok 4.6 in GitHub Copilot
- Gemini 3.7 Flash in GitHub Copilot
- GPT-5.6 Ultrafast mode