Skip to main content

OpenAI Unveils GPT-6Astra: A Leap Toward AGI with Bold Security Claims

On September 3, OpenAI pulled back the curtain on its latest flagship model, GPT-6Astra, and the company didn't mince words. President Greg Brockman declared at the press event that this release signals the industry's arrival at the "General Artificial Intelligence (AGI) era." That's a bold statement, but Astra's numbers back up some of the hype.

A Model That Solves What Others Can't

Before the public launch, the internal version of Astra had already made waves by cracking ten long-standing problems in mathematics and theoretical computer science. It also showed significant gains in agent coding and cybersecurity—so much so that it became the first OpenAI model to hit the "critical" threshold in the company's own preparedness framework for cyber risks.

Benchmark Blowout

On the ARC-AGI-3 test, Astra scored an eye-popping 98.6%, leaving competitors in the dust—GPT-5.6Sol managed just 7.8%, and Claude Opus5 trailed at 30.2%. In FrontierMath Tier4(v2), it hit 97.6%, again far ahead of the pack. Other tests tell the same story: GPQA Diamond (96%), DeepSWE v1.1 (74.1%), Terminal-Bench Science0.1 (64.6%), Agents' Last Exam (59.3%), and AutomationBench (41.4%).

Image

These aren't just incremental improvements; they're the kind of leaps that make you sit up and take notice. But raw performance is only half the story.

Security Gets a Serious Upgrade

Astra also comes with a beefed-up safety net. OpenAI says the model includes new monitoring functions designed to prevent unauthorized operations, and it can more reliably reject harmful network requests. That's a crucial step, especially as AI agents start to take on more autonomous roles.

Image

Who Gets It First?

Right now, Astra is only available to enterprise cybersecurity customers who've been granted access through OpenAI's Daybreak program. The core security features are still restricted to "trusted defenders." Over the next few days, it'll roll out to ChatGPT Plus, Pro, Business, and Enterprise subscribers.

The Price Tag

If you're hoping to try it out, be prepared to pay. Astra is priced at $10 per million input tokens and $50 per million output tokens—matching Anthropic's Claude Fable5.1. That's a premium price for a premium model, and it sets a high bar for competitors.

What This Means for the Industry

Astra's commercial deployment isn't just another product launch. It raises the threshold for what advanced AI agents can do, and it opens up a new competitive landscape where high computing power and strong security are non-negotiable. As the industry edges closer to AGI, the question isn't just who can build the smartest model—it's who can do it safely and responsibly.

Key Points

  • AGI Era Declared: OpenAI's president says GPT-6Astra marks the industry's entry into the AGI era.
  • Record Benchmarks: Scores 98.6% on ARC-AGI-3 and 97.6% on FrontierMath Tier4(v2), far outperforming rivals.
  • Cybersecurity Breakthrough: First OpenAI model to reach "critical" threshold in its preparedness framework.
  • Limited Availability: Initially only for enterprise cybersecurity customers via Daybreak program; broader rollout to ChatGPT subscribers coming soon.
  • Premium Pricing: $10 per million input tokens, $50 per million output tokens, matching Claude Fable5.1.