OpenAI unveils GPT-6 Astra, declaring the arrival of the 'AGI Era'
OpenAI has officially launched its newest frontier AI model, GPT-6 Astra, ushering in what executive leadership described as the arrival of artificial general intelligence (AGI). Trained on over 100,000 GPUs at the company’s Stargate facility in Texas, Astra represents OpenAI's largest pre-training run to date and marks a massive architectural leap in autonomous computer use, complex reasoning, and long-horizon workflows.
The model is rolling out across ChatGPT Plus, Pro, Business, and Enterprise tiers, alongside full API availability on Microsoft Foundry and OpenRouter.
Built as an Autonomous Computer Operator
Unlike previous generations designed primarily around text chat or code completions, GPT-6 Astra functions as an end-to-end computer operator. Equipped with an expansive 1.05-million-token context window, the model can natively navigate desktop and web browser environments, operate software applications, inspect screens, execute QA testing, and manage multi-step administrative workflows.
Key capability benchmarks highlighted by OpenAI include:
Computer Use & Desktop Automation: Astra leads state-of-the-art benchmarks in computer interaction, scoring 64.6% on complex operational tasks compared to 52.6% for Claude Fable 5.1. It can execute end-to-end tasks like tax preparation, job searching, and CRM updates.
Prompt-to-Artifact Generation: Through a new "Sites" feature in ChatGPT, Astra can directly build, host, and deploy functional web applications, interactive 3D environments, and playable video games, ranging from WebGL shader scenes to procedural game maps - all from single natural language prompts.
Advanced Coding and Research: Astra outperforms predecessor models across software engineering, scientific data analysis, and legal research benchmarks while consuming up to 20% fewer tokens during long-horizon coding tasks.
Novel Reasoning Technique and Safeguard Restrictions
Underneath its capabilities, Astra utilizes a novel reasoning technique termed recurrent depth. The architecture enables the model to process complex multi-step problems with heightened internal reasoning, though OpenAI noted that aspects of this chain-of-thought process are obscured from external view.
The model’s advanced autonomous capabilities prompted OpenAI to implement strict system-level safeguards before release:
Cybersecurity gatekeeping: Following internal red-teaming evaluations, OpenAI reported that Astra triggered advanced risk thresholds for cyber-offensive capabilities. As a result, public API access to Astra's most capable cybersecurity and vulnerability-exploitation features remains heavily restricted.
Alignment gains: Safety benchmarks indicate zero authorized target overreaches during automated agent evaluations—a sharp drop from the 48% boundary-drift rate seen in earlier experimental checkpoints.
Industry Impact and Enterprise Adoption
Enterprise partners, including Jane Street, Microsoft, and early academic testers, have integrated Astra to handle multi-step financial modeling, automated bug fixes in production code bases, and complex document generation adhering to strict corporate templates.
