OpenAI's GPT-6 Astra rated critical for cyber risk
At $10 per million input tokens, a perfect ExploitBench score and an August development pause: what OpenAI's new flagship changes.
Symbolic image: a server rack with blinking status lights while abstract diagrams move across wall monitors in the background.
OpenAI has launched GPT-6 Astra, the first of its own models the company classifies as critical for cybersecurity under its Preparedness Framework.
At a glance
- Pricing: $10 per million input tokens, $50 per million output tokens; Fast Mode runs 2.5x faster at double the price.
- ExploitBench: 100 percent, up from 78.5 percent for GPT-5.6 Sol.
- Internal V8 test: 39 percent of flaws found versus 5.5 percent before; two zero-days surfaced.
- Humanity's Last Exam: 57.2 percent for Astra against 65 percent for Fable 5.1, per heise online.
- FrontierMath Tier 4: 97.6 percent versus 87.8 percent for Claude.
OpenAI has released GPT-6 Astra to a small set of organizations first, with a wider rollout to ChatGPT Plus, Pro, Business and Enterprise customers and to the API expected within days. Enterprise workspaces do not get it automatically — an administrator has to switch it on.
What it costs
API access is priced at $10 per million input tokens and $50 per million output tokens. A Fast Mode trades money for latency: 2.5x the speed at twice the price, according to heise online. Pro, Business and Enterprise tiers also get a separate GPT-6 Astra Pro variant.
The first critical cyber rating
Astra is the first OpenAI model the company itself files as critical for cybersecurity under its Preparedness Framework. It scores 100 percent on ExploitBench, where GPT-5.6 Sol reached 78.5 percent. Tested internally against the V8 JavaScript engine, it surfaced 39 percent of the flaws, up from 5.5 percent, and turned up two zero-days along the way.
The rating carries weight precisely because it is self-assigned rather than imposed from outside. OpenAI had already paused the model's internal development in August over security questions it had not resolved.
AGI talk without the AGI label
President Greg Brockman framed Astra as the moment later observers will point to as the arrival of AGI, while stopping short of calling it AGI. Chief Scientist Jakub Pachocki was more guarded, noting that the more capable models become, the harder it gets to know what they can actually do. Astra's written reasoning is also harder to monitor than that of its predecessors.
Not ahead everywhere
The lead is uneven. Astra takes 57.2 percent on Humanity's Last Exam against 65 percent for Fable 5.1, and 61.2 against 65.7 on the Artificial Analysis Index. On FrontierMath Tier 4 the order flips: 97.6 percent for Astra, 87.8 percent for Claude.
What we could not check
Every figure here comes from the heise online report. OpenAI's own announcement page and a report by The Verge were unreachable for us at the time of writing, so none of these numbers — including the benchmark results and the critical rating — could be cross-read against the primary source.
FAQ
How much does GPT-6 Astra cost?
$10 per million input tokens and $50 per million output tokens, with a Fast Mode at 2.5x the speed for twice the price.
Why did OpenAI rate GPT-6 Astra critical for cybersecurity?
It reached 100 percent on ExploitBench and found 39 percent of V8 flaws in internal tests, up from 5.5 percent, including two zero-days.
Is GPT-6 Astra AGI?
No. Greg Brockman calls it the point observers may later mark as AGI's arrival, but he does not apply the label to Astra itself.
