OpenAI launches GPT-6 Astra with critical cyber capability

A year after unveiling GPT-5, OpenAI has announced the debut of GPT-6 Astra, which will commence rollout immediately. OpenAI touts Astra as the world's most intelligent and aligned model, marking a new era for the AI company. However, Astra also introduces heightened cybersecurity risks. Shortly before Astra's launch, OpenAI disclosed that the model had been classified as a critical cybersecurity risk, but assured that safeguards would be implemented for a successful rollout.

GPT-6 Astra is the first OpenAI model to receive this highest threat designation in any domain, according to the company's Preparedness Framework, which assesses risk across three areas: chemical/biological, self-improvement, and cybersecurity.

Despite the heightened cybersecurity concerns, OpenAI affirmed that GPT-6 Astra is generally safer than its recent models, particularly in high-risk scenarios. The system card for GPT-6 Astra indicates superior performance on safety evaluations, which assess risks like emotional reliance, engagement with self-harm requests, and inappropriate responses to users under 18.

Additionally, GPT-6 Astra demonstrates improved factuality and reduced hallucination rates compared to its predecessors, with fewer factual errors and hallucinations than GPT-5.6 Sol and other recent models. Benchmark scores reveal marked improvements across various tests, including Terminal-Bench Science 0.1, FrontierMath Tier 4 (v2), GPQA Diamond, Humanity's Last Exam (with tools), ExploitBench, Exploit Gym, and ARC-AGI-1.