On September 3, 2026 (US time), OpenAI announced its new flagship model, GPT-6 Astra. The company positions this model as "the most intelligent and aligned model in the world."
Model ReleasesSecurityOpenAIGPT-6 Astra
OpenAI Announces Next-Generation Model GPT-6 Astra, with Cybersecurity Capabilities Reaching "Critical"
This article is a translation. Read the Japanese original
In terms of its ability to autonomously operate PC screens, GPT-6 Astra has achieved a significant speedup, requiring approximately 40 minutes per 1 task, compared to about 75 minutes for the previous generation, GPT-5.6 Sol. It also demonstrates high performance in mathematics and science. Specifically, it recorded a score of 97.6% on the "FrontierMath Tier 4 v2," a top-tier benchmark created by researchers as a short-term research objective.
On the other hand, this model became the first to reach the "Critical" level for cybersecurity capabilities within OpenAI's "Preparedness Framework." This means that, given appropriate tools and access, the model possesses the ability to discover unknown vulnerabilities and develop methods to exploit them without human intervention.
Due to these safety concerns, the deployment immediately following the announcement was highly chaotic. Access is currently limited to organizations participating in the application-based cyber defense programs "Daybreak" and "Trusted Access." Regarding the delay in deployment to paid subscribers, OpenAI CEO Sam Altman apologized on X (formerly Twitter), describing the situation as "messy." As compensation, the company has also announced the distribution of "accumulated reset rights," which allow users to reset their usage limits after exhausting them.
Sources
- 人間なしでシステムに侵入 OpenAI初「最高危険度AI」の「Astra」は何がヤバいのか？ - ITmedia （Google News: OpenAI、2026-09-07）
- PC操作75分が40分に。GPT-6 Astraが変えたのは賢さより手数｜ZERO ─ 科学と事業のあいだ で - note （Google News: OpenAI、2026-09-04）