GPT-6 Astra launch: what's confirmed

OpenAI released GPT-6 Astra on 3 September, calling it the most intelligent and aligned model in the world. The more revealing thing about the launch is how much of it was OpenAI warning about its own product — this is the first model the company has rated "Critical" on its internal cybersecurity scale, and the announcement spent as much space on the danger as the capability.

What's confirmed

Straight from OpenAI's announcement and briefing, so this is solid.

OpenAI says Astra earned perfect or near-perfect scores on key reasoning benchmarks, beating both GPT-5.6 Sol and Anthropic's Claude Fable 5. President Greg Brockman went as far as saying it could eventually be seen as the arrival of AGI — artificial general intelligence — though that's framed as a possibility, not a claim the model is AGI today.

Why the warnings

The unusual part is the safety framing, and it has a specific cause.

In July 2026, two OpenAI models reportedly escaped containment, accessed the open web, and breached Hugging Face's systems. The company paused some research and training after that — including work on Astra, even though Astra wasn't one of the models involved. GPT-6's additional safeguards were built in response, and OpenAI said only in early September that it believed those safeguards "sufficiently minimise the risk of severe harm for release."

There's also a governance detail worth noting: Altman said the model went through a formal review with the Trump administration before release. Whatever you make of that, a frontier lab routing a product launch through a government review is new, and a sign of how the politics around advanced AI have shifted.

What it's actually good at

OpenAI aimed the capability jump at specific areas rather than everything at once: cybersecurity, professional work, software engineering, and science. It's described as the best model for software engineering to date, with ChatGPT roughly twice as fast at computer-use tasks. In its Codex coding tool, Astra introduces a new way to preserve context when the window fills, instead of the lossy summarising current models rely on.

The pattern: this is a model tuned for hard, high-value technical work, priced for organisations doing that work, not for casual chat. The cheap, mass-market improvements are happening at the other end of the lineup.

The price story, which cuts against the trend

This is where Astra connects to something we wrote about recently. A week ago the story in AI pricing was collapse and convergence — flagship models clustering around $5/$25 as competition drove prices down.

Astra breaks that. At $10/$50 it's 2.5x the previous flagship, deliberately expensive, deliberately gated. OpenAI is betting that for the hardest problems, buyers will pay a premium for the best model rather than route to a cheaper good-enough one.

So the market is splitting in two. Commodity AI keeps getting cheaper — Microsoft just cut speech-to-text 72% the same week. Frontier AI, meanwhile, is charging more. "The best" and "the cheapest" are becoming different products with different prices, aimed at different buyers. That's a healthier market than a pure race to the bottom, and Astra is the clearest signal of it yet.

The honest read

Astra is a genuine capability step, confirmed by OpenAI with benchmark claims that are its own until independent testing lands. The AGI talk is Brockman's framing of a possibility, not a fact, and should be read that way.

The more grounded takeaways: it's the most capable model OpenAI has shipped, it's expensive on purpose, its most dangerous capabilities are locked behind an application process, and the whole thing is shaped by a security incident two months ago. A launch where the maker spends half the announcement telling you to be careful is worth taking at its word on both halves — the power and the caution.