Not long ago, the OpenAI company unveiled the latest entrant in its line of AI models: GPT-6 Astra. According to its leaders, this launch marks the dawn of the era of artificial general intelligence. What infrastructure underpins this AI? What will it be used for? And why has its brief rollout sparked controversy?
A million-token contextual memory
As stated in an official OpenAI release, the firm has launched GPT-6 Astra, its most advanced AI model to date. For the company’s executives, the appearance of this new model signals the onset of the era of artificial general intelligence (AGI). This concept refers to an AI capable of rivaling a human being across the full spectrum of cognitive tasks. The model can thus understand novel contexts, reason in the face of the unexpected, and learn, without any guidance.
Unlike simple text generators, GPT-6 Astra behaves like a true agent able to directly interact with the digital environment. It can, for instance, navigate the Web, fill out forms, manipulate spreadsheets, and autonomously use sophisticated professional software such as Blender or engineering tools.
Carrying a contextual memory of one million tokens and capable of generating up to 128,000 tokens in output, the AI has passed several tests. Demos have shown that GPT-6 Astra can, for example, design a 3D scene—or code a video game—while simultaneously performing other actions, such as ordering food online. Primarily, the AI will serve IT automation, software development and advanced engineering, scientific research and high-level mathematics, as well as cybersecurity (audits).
The first AI to cross OpenAI’s defined critical risk threshold
According to OpenAI, GPT-6 Astra represents the largest training effort in OpenAI’s history. The manufacturer mobilized no fewer than 100,000 graphics processing units at the Stargate site located in Texas, United States. Furthermore, it is worth noting that OpenAI’s previous models provided a significant portion of supervision during the learning phase for the very first time. Moreover, GPT-6 Astra outperforms its predecessors, with scores of 97.6% on FrontierMath Tier 4 (Mathematics), 96.0% on the doctoral-level GPQA Diamond benchmark (Sciences), and 57.9% on Terminal-Bench (Software Engineering).
Par ailleurs, le modèle est le premier à franchir le niveau de risque critique défini par le cadre de préparation d’OpenAI. En effet, celui-ci est capable d’identifier et d’exploiter de manière autonome des vulnérabilités de sécurité complexes de type « zero-day ». Il s’agit de failles informatiques critiques totalement inconnues de l’éditeur, du fabricant du matériel ainsi que du public. Il n’est donc pas surprenant de constater que la version grand public et commerciale de GPT-6 Astra intègre de fortes restrictions. Dans les faits, celle-ci refuse de générer des cyberattaques ou des codes malveillants avancés.
Malgré ces restrictions, l’IA suscite de vives controverses et des débats intenses. Les craintes se cristallisent justement autour du franchissement du risque critique et de la perte de contrôle potentielle. En raison de ses capacités d’automatisation et de piratage, le modèle a déclenché l’alerte auprès des régulateurs en Europe. Au regard de l’AI Act européen, GPT-6 Astra est considéré comme étant un modèle à risque systémique. Ainsi, OpenAI devrait s’acquitter d’obligations de transparence extrêmement lourdes et subir une surveillance stricte des éventuels incidents à venir.
