Summary
OpenAI launched GPT-6 Astra on September 3rd, 2026, and Matt Wolfe secured early access to test it before the public rollout to ChatGPT Plus, Pro, Business, and Enterprise users. This video breaks down the key benchmarks, shares Wolfe’s personal hands-on results, and rounds up what other early access testers managed to build.
The benchmark highlights are significant: ARC-AGI-3 jumped from 7.8% to 99.9%, Terminal Bench science went from 22% to 64.6%, and Automation Bench nearly doubled from 18.1% to 41%. On the Deep Suite coding benchmark, GPT-6 Astra scores 74.1% — strong, but surprisingly close to Meta’s Muse Spark 1.3 (75.4%) and only marginally ahead of Claude Opus 5 and Gemini 3.8 Flash. Cost sits around $167 per task. Interestingly, Artificial Analysis’s aggregated benchmark places Astra in fifth place, which Wolfe finds at odds with his subjective experience of the model.
Wolfe demonstrates building a fully functional “Mega Bonk”-style game in just 8 minutes (previously a 1.5–2 hour task) and generating an interactive world simulator complete with climate sliders, population modeling, and terrain editing. He also tests a creative website prompt and walks through early community builds. The video closes with honest skepticism about benchmark reliability versus real-world feel — a recurring theme as Wolfe compares Astra favorably to Claude Opus 5 and Fable 5 despite the numbers suggesting near-parity.
📺 Source: Matt Wolfe · Published September 03, 2026
🏷️ Format: Review
