GPT-6 Astra Launch: 'Critical' AI Declares AGI Era, Split Scorecard vs Claude
GPT-6 Astra is the flagship AI model OpenAI released on September 3, 2026. For office workers, it means an AI that fills forms and works spreadsheets is coming to paid ChatGPT plans. Paused in August, it returned with the first-ever 'critical' cyber rating and Greg Brockman's line, "Welcome to the AGI era."
Why GPT-6 Astra Is Called "the Paused Model"
On August 18, OpenAI halted Astra's reinforcement learning (the trial-and-error stage of training) for two weeks as it neared its 'critical' cyber threshold, days after a July security incident. Claude Fable 5.1 (Sept 1) and Gemini 3.8 Flash (Sept 2) shipped first; Astra then cleared a voluntary US pre-release review of up to 30 days with, per Brockman, nothing to change.
GPT-6 Astra Benchmarks vs Claude Fable 5.1
OpenAI's table shows a split. Astra scored 64.6% on Terminal-Bench-Science, a science-research test, vs 52.6% for Fable 5.1, and led business automation 41.4% to 31.4%. On OSWorld, a computer-use test, one task takes about 40 minutes, half GPT-5.6 Sol's 75. Yet Fable 5.1 leads the Artificial Analysis Intelligence Index 65.8 to 61.3 and Humanity's Last Exam with tools 65.0 to 57.2. ARC-AGI-3: 99.9% on OpenAI's harness, 62.7% on ARC Prize's standard one.
What a 'Critical' Cyber Rating Means for the AI Industry
Astra scored 100% on ExploitBench and found two zero-day vulnerabilities in testing, so consumer tiers refuse advanced cyber tasks and offense stays with vetted defenders (Daybreak). With the US review and the EU AI Act's Article 92 powers, live since August 2, a frontier launch now resembles an inspection. Pricing matches Fable 5.1 at $10 in and $50 out per million tokens, with a 100,000-GPU Stargate run in Texas behind that price.
Is This AGI? What to Watch Next
Brockman left the AGI call to readers but said, "I think we're there." With Claude ahead on the overall index, "best model" depends on the test. A computer-operating model reaches paid ChatGPT tiers, the API, and AWS in days. How Anthropic and Google answer, and how fast regulators move on 'critical' models, are the questions now.
Key Takeaways
① A 16-day turnaround - paused Aug 18, shipped Sept 3 as the first 'critical' model; Brockman: "AGI era"
② Split scorecard - beats Claude on science 64.6% vs 52.6%, trails the index 61.3 vs 65.8
③ Safety-permit race - offense limited to vetted defenders; US review and EU Article 92 turn launches into permits
The AI race is moving from score tables to who can release safely. Watch how fast computer-operating AI reaches the workplace, not the label.
👉 OpenAI Pauses Astra After First-Ever Critical Cyber Risk Flag (2026) - also worth a read.
📌 Sources: OpenAI, Axios, Fortune, The New Stack, The Next Web, ARC Prize, Google (2026)
댓글
댓글 쓰기