OpenAI has started granting access to GPT-6 Astra, its latest and most powerful AI model to date. Astra demonstrates exceptional capabilities across complex tasks including interactive application use, advanced mathematics, and cybersecurity vulnerability detection, positioning it as a major step toward artificial general intelligence.
- Scores 98% on complex math benchmarks and excels in AGI-related tests
- Identified as a critical cybersecurity risk, leading to enhanced safety guardrails
- Available initially for cybersecurity research with $1B in support credits
What happened
OpenAI Group has officially started opening access to GPT-6 Astra, its latest large language model boasting state-of-the-art performance across a range of demanding tasks. The model excels in coding, browsing, and interacting with graphical user interfaces, representing a major advance in AI capabilities. Prior to its public release, Astra already solved decades-old mathematical problems and made breakthroughs in computational complexity theory.
The model completed some of the most challenging AI benchmarks with near-perfect scores, including FrontierMath Tier 4, ARC-AGI-3, and ExploitBench. Particularly notable is its 99.9% score on an AGI-focused benchmark, reflecting its ability to learn new tasks quickly. Due to Astra’s demonstrated ability to autonomously hack numerous secure systems, OpenAI classified it as a critical risk and delayed its launch to build comprehensive guardrails.
Why it matters
GPT-6 Astra sets a new standard for AI performance by combining advanced reasoning skills with highly sophisticated cybersecurity capabilities. Its success in tackling a range of complex mathematical and computer science challenges pushes forward the pursuit of artificial general intelligence (AGI), a key milestone in AI development. The model’s capacity to find and exploit software vulnerabilities autonomously poses both powerful opportunities and new risks.
The cybersecurity risk designation underscores the need for robust safety frameworks in advanced AI deployments. OpenAI’s decision to delay Astra’s release while implementing guardrails demonstrates a responsible approach to managing these risks. Astra’s innovative data handling techniques—compressing prompts without losing original context—enhance output accuracy and usability, setting a precedent for future large language models.
What to watch next
Initially available through the Daybreak program, Astra is targeted at organizations conducting cybersecurity research, particularly essential service operators like government agencies and utilities. OpenAI has committed $1 billion in credits for model usage, training, and support to help these groups enhance their threat defense capabilities. Collaboration with nonprofit organizations such as MS-ISAC will further support public sector security initiatives.
Over the coming days, OpenAI plans to expand Astra’s availability by integrating it into ChatGPT, Codex, and its API services. Observers should monitor how Astra’s rollout influences the balance of AI innovation and security risk, and how its advanced learning capabilities impact development toward AGI. The effectiveness of safety measures applied around Astra will also be a critical focal point for regulators and the broader AI community.