OpenAI's GPT-6 Astra Is Shockingly Good at Almost Everything — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI released GPT-6 Astra on September 3 at $10 per million input tokens and $50 per million output—2.5x the rate of GPT-5.6 Sol, per Artificial Analysis
- GPT-6 Astra hit 72.6% on the OSWorld 2.0 desktop-agent benchmark in roughly 40 minutes per task, versus 65.7% in 75 minutes for Sol, with OpenAI's headline feature being direct mouse-and-keyboard computer use
- GPT-6 Astra is the first model OpenAI has rated at the critical cybersecurity threshold, meaning it can autonomously find unknown software flaws and build working attacks without human guidance
- OpenAI president Greg Brockman used the launch briefing to declare the arrival of AGI
- Early-access developers built a street-by-street Manhattan replica in Unreal Engine, a browser-based 3D Hangzhou in 24 minutes via Three.js, and a 12-person multiplayer shooter called Astral War in a single day
- GPT-6 Astra posted the best score yet on Auggie's Bach chorale benchmark, with no voice-leading errors and the first-ever passing tones recorded on the test
- GPT-6 Astra dropped roughly 80 Elo points on Artificial Analysis's benchmark of economically valuable professional writing and landed 11th (1995 Elo) on Louis-François Bouchard's internal writing benchmark, versus 6th for its predecessor
Why it matters: The 2.5x price hike over GPT-5.6 Sol for a model sharper at spatial and agentic tasks but measurably worse at writing means customers are paying more for a specialized upgrade, not a broad one. OpenAI rating Astra at the critical cybersecurity threshold also raises the stakes on guardrails, since the same model now driving a mouse can also autonomously discover and weaponize software flaws.
Ask SkimNews
