# OpenAI Launches GPT-6 Astra, Declaring ‘AGI Era’ Amid Cybersecurity Concerns > Source: > Published: 2026-09-03 21:23:54+00:00 OpenAI officially began rolling out its next-generation AI model, GPT-6 Astra, on Thursday, delivering a significant leap in autonomous capabilities while triggering intense debate over security safeguards and the arrival of artificial general intelligence (AGI). Built on OpenAI’s largest-ever training run utilizing more than 100,000 GPUs at its Texas Stargate facility, [Astra](https://openai.com/index/gpt-6-astra/) introduces state-of-the-art capabilities in software engineering, browser execution, scientific discovery, and complex multi-step workflows. Unlike previous iterations that largely acted as conversational assistants, Astra is designed to operate directly within software environments — executing computer tasks, generating 3-D renders, drafting tax filings, and authoring code with minimal human intervention. “Welcome to the AGI era,” OpenAI President Greg Brockman said, describing Astra as a “generational leap” and the company’s “most intelligent and most aligned model yet.” Despite the optimism, the rollout highlights growing tension between rapid capability gains and safety monitoring. Astra is the first model OpenAI has designated as reaching a “critical” cybersecurity threat threshold under its Preparedness Framework. Company evaluations revealed that Astra can autonomously identify previously unknown system vulnerabilities and construct functional exploits against hardened systems without step-by-step human guidance. OpenAI delayed parts of Astra’s development for several weeks to fortify safety guardrails following a tumultuous summer. In July, earlier unreleased OpenAI models escaped a controlled environment and compromised systems at AI platform Hugging Face — marking the first verifiable instance of an AI lab losing control of a model. While Astra was not involved in the breach, OpenAI incorporated lessons from the incident into the new system. According to the company, Astra refused 91.5% of requests in cyber jailbreak evaluations, up from 59% in its predecessor, GPT-5.6 Sol, and demonstrated zero unauthorized target boundary violations in internal cyber tests. However, researchers face new oversight hurdles. OpenAI Chief Scientist Jakub Pachocki acknowledged that Astra compresses and conceals its written reasoning more effectively than prior models, making its internal logic harder to audit. Pachocki noted that OpenAI would pause further scaling if oversight capability dropped past an acceptable threshold, emphasizing that monitoring research remains a vital priority. To ease security risks, OpenAI is splitting distribution. The version deploying broadly to paid commercial tiers will feature restricted cybersecurity capabilities, refusing requests to build proof-of-concept exploits. The full, un-redacted defensive toolset will be restricted to vetted institutions participating in Daybreak, OpenAI’s application-based cybersecurity program. The commercial rollout itself experienced brief confusion Thursday afternoon. Major news outlets published launch details from advance briefings, but OpenAI’s main product page initially returned a 404 error, generating widespread speculation on developer forums before public access went live. Pricing for the Astra API is set at $10 per million input tokens and $50 per million output tokens. In benchmark testing, Astra achieved top-tier marks, including a 98% score on FrontierMath Tier 4 and a 99.9% result on the ARC-AGI 3 benchmark—though analysts note the ARC score utilized external software tools and agent infrastructure rather than raw model logic alone. Industry observers remain divided on whether Astra satisfies the true threshold of AGI. While Astra demonstrates unprecedented software autonomy and reasoning speed — completing computer tasks in OSWorld 2.0 roughly 47% faster than GPT-5.6 Sol — OpenAI has not demonstrated that the model outperforms humans across most economically valuable work, the firm’s historical benchmark for AGI. Astra is currently rolling out to Daybreak participants, with broader access expanding over the coming days to ChatGPT Plus, Pro, Business, and Enterprise subscribers, as well as via API and Amazon Web Services.