OpenAI Astra Trains on 100,000 GPUs and Targets AGI

OpenAI’s Astra has been presented as a major step toward artificial general intelligence (AGI). The company says Astra is its first model trained on more than 100,000 GPUs and can operate computer software directly, potentially reducing the need for separate APIs across applications. OpenAI first publicly identified Astra in August 2026, after the reported release of GPT-5.6. Later claims said Astra solved 10 advanced mathematics and theoretical computer science problems with verified proofs at an estimated cost of about $2,000. On 1 September, it reportedly reached the “Critical” cybersecurity level under OpenAI’s Preparedness Framework, including the ability to find and exploit previously unknown vulnerabilities. Access to this capability remains restricted. OpenAI president Greg Brockman said Astra has been used to scan the company’s systems and help develop security patches. However, OpenAI also acknowledged weaknesses in its sandbox defences after a Hugging Face security incident. Brockman said capability, safety and alignment must advance together, and suggested Astra or a successor could meet commonly accepted AGI standards. A Playco case study reported 50% fewer manual fixes during game prototyping. OpenAI is also developing its Jalapeño chip with AI-assisted design while maintaining plans to deepen its Nvidia partnership. These developments could support long-term demand for GPUs, semiconductors, data centres and cybersecurity services. However, Astra’s capabilities, commercial impact and AGI trajectory still require independent verification. Astra does not directly change cryptocurrency fundamentals, so any crypto-market reaction is likely to be indirect, sentiment-driven and limited in the short term.