AI
Inside OpenAI's GPT-6 Astra: Why the AGI Era Is Now on the Table
Something shifted last week, and the whole industry felt it. On September 4, OpenAI began rolling out GPT-6 Astra, which the company calls its most intelligent and most aligned model yet. The benchmark sheet reads like a dispatch from the future: 98 percent on FrontierMath Tier 4, a set of 50 math challenges that occupy human mathematicians for weeks, 99.9 percent on ARC-AGI-3, which measures how fast an AI learns brand new tasks, and a perfect 100 percent on ExploitBench. It also posted 96 percent on GPQA Diamond, the graduate level science gauntlet, and set new records on agentic tests like Terminal-Bench 4.0.
The numbers were foreshadowed before launch. In its final testing phase, Astra solved long standing Erdos problems, math puzzles that have waited decades for answers, and pushed forward computational complexity theory. Its signature skill is computer use. In OpenAI's launch demo it filled out a tax return from a W-2, drafted a legal contract, searched apartment listings, and built a 3D game. It scores 72.6 percent on OSWorld 2.0, a real desktop task benchmark, finishing jobs about 47 percent faster than its predecessor. Developers get better economics too: on DeepSWE v1.1, Astra's top configuration cuts estimated API cost per completed task by roughly 57 percent.
Capability at this level arrives with serious responsibility, and OpenAI treated it that way. Astra is the first OpenAI model to reach “critical” under the company's safety framework for cybersecurity, recognizing how quickly it spots software vulnerabilities. Engineers spent several additional weeks hardening guardrails before release. Exploit creation stays gated at launch: Astra helps with secure code review and patching, while proof of concept exploits wait for the Daybreak research program. OpenAI paired the launch with $1 billion in credits to help government agencies, utilities, and essential service operators defend their networks, with training support from the nonprofit MS-ISAC.
The milestone electrified the industry. NVIDIA chief Jensen Huang posted congratulations on X, noting Astra trained on about 100,000 NVIDIA Grace Blackwell systems and declaring that AGI has arrived, with 400,000 more GPUs coming online next. OpenAI president Greg Brockman called the launch a turning point, saying it feels reasonable to describe this as the AGI era. Independent testers add useful context: on some broad third party composites, rival models still hold the lead, which shows how rapidly the entire field is advancing.
The practical upside lands directly in your hands. Tools that once demanded teams of analysts, engineers, and mathematicians now run on a single screen, and access is expanding in waves from select organizations to consumer and enterprise tiers. A freelancer can delegate research workflows, a small business can automate its paperwork, and a student can get one on one tutoring at the highest level. The next era of building belongs to builders everywhere, and the pace keeps accelerating.
Sources
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.