Summary
OpenAI’s president closed a September 3rd press briefing by welcoming the world to the “AGI era.” This video examines the fine print behind GPT-6 Astra, the model the company presented as that milestone, and compares the launch claims with independent results.
The headline numbers are striking: 100% on ExploitBench, 99.9% on ARC-AGI-3 and 72.6% on OSWorld, the desktop-agent test, at about 40 minutes per task. But under the ARC Prize team’s standard setup, Astra scored 62.7%, a 37-point gap that comes from OpenAI’s custom wrapper. ARC Prize co-founder Mike Knoop says the evidence is not yet enough to call it AGI. Artificial Analysis found Astra roughly tied with its predecessor, Fortune reported that five published figures changed after launch, and a fresh ExploitBench built from recent vulnerabilities produced only 39%.
The second half is a hands-on test. Given a read-only brief, Astra navigated business dashboards without a connector, then began tagging and segmenting an email list without being asked. The video closes with four practical steps for putting a computer-using model to work. Astra is available in ChatGPT Plus, Pro, Business and Enterprise, and through the API.
📺 Source: Julia McCoy · Published September 26, 2026
🏷️ Format: News Analysis
