Super Data Science: ML & AI Podcast with Jon Krohn 1026: OpenAI’s GPT-6 Astra
34 snips
Sep 11, 2026 A deep dive into GPT-6 Astra’s leap in computer use, coding, abstract reasoning, scientific research, and professional workflows. The discussion explores its ability to handle ambiguity and corrections, plus the safety risks raised by powerful cybersecurity capabilities. It also weighs whether Astra signals AGI or simply marks another major step toward it.
AI Snips
Chapters
Transcript
Episode notes
Astra Scales Training And Pricing
- GPT-6 Astra is OpenAI’s largest training effort, using more than 100,000 GPUs and model-generated supervision.
- API access costs $10 per million input tokens and $50 per million output tokens, with a faster mode costing twice as much.
Astra Makes Computer Agents Faster
- Astra improves desktop-agent accuracy while sharply reducing task time, completing OS World tasks in 40 minutes versus Sol’s 75.
- ScreenSpot Pro shows a larger jump, reaching about 93% compared with Sol’s 77%.
Astra Preserves Coding Context
- Astra’s TerminalBench score rises to 58% from Sol’s 37%, roughly matching Claude Fable 5.1.
- Its Codex sessions preserve searchable notes across context windows instead of compressing history into one summary.