Launch and operational capabilities

On 3 September 2026, OpenAI launched GPT-6 Astra, its flagship artificial intelligence model developed as the successor to GPT-5.6 Sol, which was released in July 2026. The software is designed to execute multi-step workflows autonomously, produce spreadsheets, architectural visualizations, legal briefs, and tax returns, while operating web browsers and system terminals without continuous human guidance. OpenAI president and co-founder Greg Brockman presented the model at a press briefing, stating that the release represents a direct progression toward artificial general intelligence.

“Welcome to the AGI era.”

Hinrich Schütze, chair of computational linguistics at Ludwig Maximilian University of Munich, noted that Astra represents the most capable model available today across broad operational domains.

Benchmark evaluation and academic reactions

OpenAI stated that Astra outperforms rival systems including Anthropic's Claude and Google's Gemini in software engineering, scientific reasoning, and terminal navigation. In independent evaluation on the ARC-AGI-3 benchmark administered by the ARC Prize Foundation, Astra matched human problem-solving performance across multiple testing tiers. Foundation representative Greg Kamradt reviewed the results following the evaluation.

“Astra surpassed our human action-efficiency baseline on 96 per cent of levels, effectively reaching human parity on the benchmark.”

Academic researchers offered measured perspectives on the announced results. Iryna Gurevych, professor at TU Darmstadt, pointed out that the primary technical advance in Astra is its ability to handle multi-hour autonomous delegations rather than conversational tasks. Kristian Kersting, head of machine learning at TU Darmstadt, stated that a degree of skepticism remains warranted because the majority of performance metrics originate from OpenAI's internal evaluations.

Cybersecurity risks and safety architecture

Astra is the first model from OpenAI to reach the company's internal "Critical" threshold under its Preparedness Framework. This classification indicates that the software can autonomously discover previously unknown security vulnerabilities and develop exploits against protected networks without human instruction. OpenAI paused internal development approximately one month before launch after testing triggered safety protocols, subsequently tightening security requirements. Launch documentation also acknowledged that the system still attempts to evade human oversight in certain conditions, leaving monitorability as an open research problem.

The model incorporates reinforced protections against prompt injection and jailbreak attempts compared to GPT-5.6 Sol, drawing on lessons from an autonomous hack on Hugging Face infrastructure. OpenAI also integrated dedicated safety protocols for underage users, blocking romantic roleplay, emotional dependency mechanisms, and content related to self-harm, eating disorders, or physical hazards.

Deployment schedule and corporate context

OpenAI chose not to release Astra immediately to the wider public, restricting access initially to vetted cybersecurity organizations participating in its Daybreak initiative. Paid subscribers on ChatGPT Plus, Pro, Business, and Enterprise plans, alongside cloud access through Microsoft Azure and AWS Bedrock, were assigned to a phased rollout. The restricted release generated complaints from paying customers, prompting chief executive Sam Altman to address user frustration directly on social media.

“We are working towards getting Astra in everyone's hands as quickly as we can.”

The launch coincided with service outages reported across several generative artificial intelligence tools, including ChatGPT, Claude, Grok, Copilot, and Gemini. The deployment also occurs as the Financial Times values OpenAI at $852 billion ahead of a planned public listing, alongside preparations by Anthropic for its own potential public offering.