OpenAI Drops GPT-6 Astra: It Can Actually Hack You Now

GPT-6 Astra is built for multi-step autonomy, moving beyond simple chat to navigating complex, long-running tasks. Instead of just answering a prompt, the model now strings together entire chains of logic and tool usage to reach a goal. For now, this is a restricted rollout, focusing on security pros who are likely sweating as they watch the model score a perfect 100% on ExploitBench.

The company officially labeled this release as Critical under its Preparedness Framework, which is tech-speak for 'this thing can find and exploit unknown vulnerabilities without human guidance.' To stop the AI from going rogue, OpenAI had to implement a new monitoring layer that watches the model's 'reasoning' in real-time, occasionally killing its own processes if they look suspicious.

The safety paranoia reached such levels that OpenAI actually hit the brakes on training for two weeks after a security incident, shuffling infrastructure and tightening the screws on alignment before resuming work. Even the math skills got a glow-up, with the model proving complex theoretical problems using Lean 4—a formal language that actually checks if the math is correct, rather than just hallucinating a pretty answer.

As for the pricing, it is a premium affair at $10 for input and $50 for output per million tokens, making it significantly pricier than GPT-5.6 Sol. Greg Brockman finished the reveal by declaring the dawn of the AGI era, though the definition of 'universal intelligence' seems to shift whenever the invoice arrives. We are officially at a point where the cost of controlling an AI is starting to rival the cost of training it, creating a fun paradox where the smartest machines are also the most heavily house-arrested by their own creators.