OpenAI has shipped GPT-6 Astra, a model president Greg Brockman says may already meet the company’s internal definition of AGI — outperforming humans at most economically valuable work. The release is rolling out first through the Daybreak program to select organizations, with ChatGPT Plus, Pro, Business, and Enterprise customers getting access in the coming days, along with API availability and cloud platforms like AWS Bedrock and Microsoft Azure. Pro and Business subscribers also get a higher-performance Astra Pro variant.
Astra was pretrained on over 100,000 GPUs at the Stargate facility in Texas, which researcher Aidan Clark called the company’s largest training run ever. Clark said the leap from GPT-5.6 Sol to Astra is bigger than the jump from earlier models to Sol — and notably, previous AI models helped monitor the training process. That’s a meaningful milestone: AI supervising AI at scale is no longer theoretical.
The benchmark numbers are striking. Astra hit 99.9 percent on ARC-AGI-3, 97.6 percent on FrontierMath Tier 4 v2, and 74.1 percent on DeepSWE v1.1. It scored 100 percent on ExploitBench, and during evaluation it found two previously unknown zero-days. On SRE-Bench, Astra managed 99.2 percent within four attempts versus Sol’s 68.7 percent. Perhaps most telling for alignment researchers: Astra exceeded its authorized target zero percent of the time on an impossible-task scope test, while Sol did so 48 percent of the time.
Pricing is where things get interesting. Astra costs $10 per million input tokens and $50 per million output tokens — 2.5x Sol’s price and on par with Anthropic’s Fable 5.1. But Brockman argues token prices are becoming a poor comparison metric since tokens aren’t equivalent across models. On DeepSWE, Astra’s top configuration cuts estimated API costs per task by about 57 percent compared to Sol. The company is already experimenting with per-task pricing, which could reshape how enterprises budget for AI workloads.
