“Welcome to the AGI era” - that's how OpenAI President Greg Brockman ended yesterday's GPT-6 Astra briefing.

Marketing? Maybe :)

But some of the numbers behind Astra are difficult to ignore.

It is OpenAI's first model trained on 100,000+ GPUs, their largest training run ever.

OpenAI reports:

- 99.9% on ARC-AGI-3

- 98% on FrontierMath Tier 4

- 100% on ExploitBench

But at Intspirit, one completely different number caught our attention: 48% → 0%.

A few days ago, our CPO wrote about the OpenAI agent incident where agents broke containment, coordinated with each other and eventually attacked Hugging Face. Full story is here.

OpenAI apparently learned quite a lot from it.

They created a new evaluation inspired by that exact incident: what happens when an AI agent faces a difficult or impossible task?

Without production safeguards, GPT-5.6 Sol went beyond its authorized target in 48% of cases.

GPT-6 Astra: 0%.

And here's the paradox.

Astra is not becoming safer because it is less capable.

Quite the opposite.

It is the first OpenAI model to reach the Critical cybersecurity capability threshold.

With the right tools and access, OpenAI says Astra can discover previously unknown security vulnerabilities and develop new ways to exploit well-protected systems - without a human guiding every step!

So we now have a model that is simultaneously:

more capable → more autonomous → potentially more dangerous → and apparently much better at understanding where it should stop.

And reliability improved significantly too!!!

The chart below comes directly from OpenAI's Astra System Card. On a dataset specifically built from ChatGPT conversations previously flagged by users for factual errors, Astra shows a dramatic reduction in hallucinations compared with GPT-5.6 Sol.

This may be much more important than another +5% on a coding benchmark.

Because as we move from chatbots to autonomous agents, intelligence isn't the only metric that matters anymore.

Judgment and boundaries become part of the benchmark too.

Astra is rolling out to paid ChatGPT users and API developers over the coming days.

We'll definitely test it at Intspirit and maybe Greg Brockman is right.

Maybe in a few years we'll look back at September 2026 and say: this was the beginning of the AGI era.

Or maybe we'll laugh at that sentence :)

Either way, things just got much more interesting!

Sources:

https://openai.com/index/gpt-6-astra/?utm_source

https://deploymentsafety.openai.com/gpt-6-astra/safety-overview-gpt-6-astra?utm_source