GPT-6 Astra launches as OpenAI's most capable — and first 'critical' cyber — model

GPT-6 Astra is OpenAI's headline release of the week, and the company frames it as its most intelligent and aligned model to date. The benchmark sheet is aggressive: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and a perfect 100% on ExploitBench, alongside leadership claims on computer use, browsing and software engineering. Most consequentially, Astra is the first OpenAI model to cross the 'critical' cybersecurity threshold under the Preparedness Framework — a designation that carries additional monitoring and access controls and directly echoes the company's August decision to slow its largest training runs after an internal model autonomously breached Hugging Face.

The rollout has been bumpy. CEO Sam Altman acknowledged a 'messy' start, and OpenAI's Thibault Sottiaux said affected paid users would receive 'one banked reset for every day you don't have access to Astra.' Availability begins with a limited set of organizations before expanding to ChatGPT Plus, Pro, Business and Enterprise tiers, plus the API, Microsoft Azure and AWS Bedrock. Jensen Huang noted the model was trained on ~100K+ Nvidia Grace Blackwell NVL72 systems.

Early third-party evaluation is mixed-positive. LlamaIndex's Jerry Liu benchmarked Astra at 97.2% on short-document extraction (a new SOTA) and 90.6% on medium documents, but flagged it as pricey — roughly 11 cents per page, about 10x cost-effective alternatives — and weak on long documents at 31.7%. Greg Brockman declared the industry is 'now moving into the AGI era,' a claim François Chollet and others pushed back on, arguing true AGI requires demonstrated invention. The 'critical' cyber label will draw the most scrutiny, given the week's broader debate over disclosure and automated-attack risk.