GPT-6 Astra is the first model making OpenAI willing to declare the "AGI era"

Details

OpenAI has released GPT-6 Astra, its largest training run ever on more than 100,000 GPUs at the Stargate facility in Texas. President Greg Brockman says Astra may already qualify as AGI — or is at least within reach — under OpenAI's own definition: AI that outperforms humans at most economically valuable tasks. Researcher Aidan Clark noted the capability jump from GPT-5.6 Sol to Astra exceeds previous generational leaps, partly because earlier AI models helped monitor training. Benchmark results are sweeping. Astra scores 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 v2, 96% on GPQA Diamond (expert knowledge), 95.9% on BenchCAD, 74.1% on DeepSWE v1.1 (software engineering), and 100% on ExploitBench (cybersecurity). All scores outperform Anthropic's Fable 5 and Fable 5.1 models across most categories. Pricing is 2.5x higher than Sol's token cost, on par with Anthropic Fable 5.1. OpenAI argues per-task cost is lower for reasoning-heavy workloads given efficiency gains. Rollout begins with Daybreak program organizations, then extends to ChatGPT Plus, Pro, Business, and Enterprise customers and the API via AWS Bedrock and Azure. The safety dimension is significant: Astra is the first OpenAI model rated "critical" under the company's own safety framework. During internal testing it independently identified two previously unknown zero-day security vulnerabilities, raising dual-use risk concerns at the frontier even as OpenAI celebrates the capability milestone.