GPT-6 Astra: OpenAI’s Most Powerful Model – and the Fine Print Behind Its 99.9%

On 3 September 2026, OpenAI launched GPT-6 Astra — its most powerful model yet, and the first it trained on more than 100,000 GPUs. The headlines screamed “99.9% — AGI is here.” The reality is more interesting: Astra is a genuine leap and its biggest number comes with an asterisk. Here’s what’s real, what’s hype, and why it matters.

What GPT-6 Astra actually is

Astra is OpenAI’s new flagship, rolled out in phases — first to companies in its cybersecurity program, then to ChatGPT Plus, Pro, Business and Enterprise users, the API, and AWS. OpenAI’s VP of research Aidan Clark said it was “by far” their largest training run — the first time they pretrained on over 100,000 GPUs, at the Stargate site in Texas. It’s notably strong at using a computer, writing code, and — the part that made OpenAI cautious — cybersecurity tasks.

A first: the “critical” cybersecurity threshold

Just before launch, OpenAI said Astra is its first model to cross the “critical” cybersecurity threshold on its Preparedness Framework — the top risk level, reserved for AI that could “devise entirely new ways to cause serious harm.” VP Amelia Glaese told Reuters that, given the right tools and access, Astra can find previously unknown vulnerabilities and build working exploits across multiple well-protected systems — without human guidance at each step.

From OpenAI’s own testing:

• 100% on ExploitBench (building exploits from known flaws) — and in a fresh internal test it discovered two brand-new “zero-day” vulnerabilities, now being reported to the vendors.
• In expert assessments it built a browser sandbox-escape and an operating-system privilege-escalation chain — from an ordinary user all the way to “root.”
• On the safety side, Astra refuses 91.5% of malicious cyber requests (up from 59% in the previous model), and in “honeypot” traps it made zero unauthorized break-in attempts (vs 56% before) — OpenAI calls it its “most aligned model to date.”

Access to the most advanced cyber features is gated — starting with a small tester group and expanding through a defenders-first program called “Daybreak Blue.”

The “99.9%” needs an asterisk

This is the number everyone quoted — and where careful reading matters. On the ARC-AGI-3 reasoning benchmark, OpenAI reported 98.6% (up from just 7.8% six months earlier). But independent testing found the score depends heavily on how you run it:

So the marketed ~99.9% comes from a setup that preserves the model’s reasoning state; on a standard shared test it lands around 62.7%. As one analysis put it: the score “looked like AGI — then researchers read the fine print.” Still a real jump — just not the finish line.

The controversy: harder to watch

Astra uses a technique OpenAI calls “opaque recurrence” — it can reason using fewer (or no) visible language tokens. That makes it faster, but also makes its “chain of thought” harder for humans to monitor. As models get more capable, OpenAI admits, keeping them interpretable gets harder — a live safety debate across the whole field.

Why this matters for you

• AI capability is accelerating — coding, cyber and computer-use are jumping fast. If you’re a student, AI fluency is no longer optional.
• Benchmark literacy is a skill. The “99.9% vs 62.7%” gap is the perfect lesson: always ask how a number was measured before you believe it.
• The AGI debate is now mainstream — but “start of AGI” is a claim, not a settled fact. Stay curious and skeptical in equal measure.

FAQ