{
  "id": 5390817,
  "title": "OpenAI unleashes Astra, its most capable and controversial model yet",
  "url": "https://urgent.news/2026/09/03/openai-unleashes-astra-its-most-capable-and-controversial-model-yet",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-03T18:37:35.000Z",
  "source": {
    "name": "Fast Company",
    "slug": "fast-company",
    "url": "https://www.fastcompany.com/91601838/openai-unleashes-astra-its-most-capable-and-controversial-model-yet"
  },
  "original_language": "en",
  "account": "OpenAI has unveiled Astra, its most advanced and contentious artificial general intelligence (AGI) model yet, on Thursday. AGI refers to highly autonomous systems that surpass human performance in most economically valuable work. Astra is described as the safest model to date, developed through improvements in pre-training, reinforcement learning, and safety measures. Over 100,000 GPUs were utilized for its training. The model outperformed other leading models in various benchmark tests, including computer use, browser use, software engineering, cybersecurity, science, and professional work.\n\nAstra achieved impressive scores, such as 98.6% on ARC-AGI 3, a game-style reasoning test designed to limit advance preparation. It was rolled out to enterprise customers, including cybersecurity professionals in OpenAI's restricted-access Daybreak program, and will be available to Plus, Pro, Business, and Enterprise users, the OpenAI API, and Amazon Web Services in the coming days.\n\nAstra can perform various tasks, such as filling out forms, updating CRM records, conducting research, drafting summaries, analyzing data, building websites, running frontend QA checks, and troubleshooting problems on screen. It outperformed its predecessor, GPT-5.6 Sol, in desktop tasks and completed them 1.9 times faster. The model excels in following existing templates and producing slides, documents, and spreadsheets that match a user's writing and visual style. Astra scored 74.1% on DeepSWE v1.1 and performed above Claude Fable 5.1 while using fewer output tokens.\n\nIn terms of cybersecurity, Astra scored 100% on ExploitBench, which tests a model's ability to identify and exploit software vulnerabilities without human guidance. It was the first time OpenAI made such a claim about one of its models, raising concerns within the cybersecurity community. Astra can find unknown vulnerabilities and figure out how to exploit them across various well-defended systems. During evaluations, it discovered two previously unknown security flaws and chained them together to gain access to a software platform, demonstrating capabilities that have alarmed security professionals.\n\nOpenAI acknowledged the concerns and is taking steps to mitigate potential risks. Astra uses a technique called \"recurrent depth\" or \"opaque recurrence,\" which involves more computations happening in the model's internal, or latent, state. This allows for higher performance while making the model's written reasoning harder to monitor. OpenAI is deploying misalignment monitoring in production, using classifiers to stop potentially unauthorized activity and can also pause legitimate work if needed. The company believes Astra is less likely to go rogue than earlier models, as it refused advanced tasks such as exploit discovery in recent evaluations.",
  "summary": "“Welcome to the AGI era,” said OpenAI president Greg Brockman at the tail end of a call with reporters about the Thursday release of the company’s newest model, GPT-6 Astra. OpenAI describes AGI, or artificial general intelligence , as “highly autonomous systems that outperform humans at most economically valuable work.” The company calls Astra its most intelligent and safest model to date, built…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 6,
    "also_reported_by": [
      {
        "outlet": "Fortune",
        "title": "Safety experts warn novel design of OpenAI’s Astra model could make future AI agents harder to monitor",
        "url": "https://urgent.news/2026/09/03/safety-experts-warn-novel-design-of-openais-astra-model-could-make",
        "published": "2026-09-03T17:51:23.000Z"
      },
      {
        "outlet": "The Verge",
        "title": "OpenAI’s next big AI model has ‘entered the AGI era’",
        "url": "https://urgent.news/2026/09/03/openais-next-big-ai-model-has-entered-the-agi-era",
        "published": "2026-09-03T18:00:00.000Z"
      },
      {
        "outlet": "TechRadar",
        "title": "GPT-6 is here, but what if we just said 'no thanks' to Astra, a model so powerful that we may never fully understand it",
        "url": "https://urgent.news/2026/09/03/gpt-6-is-here-but-what-if-we-just-said-no-thanks-to-astra-a-model-so",
        "published": "2026-09-03T19:49:53.000Z"
      },
      {
        "outlet": "Techmeme",
        "title": "OpenAI commits $1B in subsidized model access, training, support, and partnerships to a new initiative aimed at protecting essential services around the world (Sam Sabin/Axios)",
        "url": "https://urgent.news/2026/09/03/openai-commits-1b-in-subsidized-model-access-training-support-and",
        "published": "2026-09-03T20:40:01.000Z"
      },
      {
        "outlet": "9to5Google",
        "title": "OpenAI launches GPT-6 Astra, ‘the world’s most intelligent and aligned model’ that you can’t use just yet",
        "url": "https://urgent.news/2026/09/03/openai-launches-gpt-6-astra-the-worlds-most-intelligent-and-aligned",
        "published": "2026-09-03T20:50:00.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}