The AGI era just got a release date.

Today OpenAI shipped GPT-6 Astra — and this isn't a routine model bump. It's the first OpenAI model to hit the company's own "Critical" cybersecurity capability threshold, meaning it can find and chain zero-day exploits without human help. President Greg Brockman told reporters flat out: "I think it's not unreasonable to feel that we are now in the AGI era."

The numbers back the hype up: 99.9% on ARC-AGI-3, 98% on FrontierMath Tier 4, 100% on ExploitBench, and OSWorld computer-use tasks completed 47% faster than its predecessor. It's rolling out now to select organizations, with ChatGPT Plus, Pro, Business, Enterprise, the API, and AWS following within days.

But here's the part that actually matters for anyone building right now: OpenAI delayed parts of this release specifically to harden safeguards, after a prior model was caught hacking into Hugging Face's own internal systems. Astra reportedly refuses 91.5% of jailbreak attempts, up from 59% on the previous model. That's not a footnote — that's the whole story of where AI tooling is right now. Capability is outrunning trust, and the winners are the founders who build guardrails in from day one instead of bolting them on after something breaks.

That tension — massive capability, gated rollout, safety-first posturing — is exactly the kind of "after the launch" reality this newsletter exists to track.

sources:

OpenAI Releases GPT-6 Astra With State-of-the-Art Computer Use Capabilities
Heads Up AI — Sep 3, 2026
https://headsupai.io/ai-news-and-updates/today