SILICON VALLEY — OpenAI released Astra, a model it defines as human-level artificial general intelligence, on the 3rd. The company said the model outperforms Claude, built by rival Anthropic, while acknowledging that security concerns grow in proportion.
Unveiling "GPT-6 Astra" on the 3rd, OpenAI said the model "boasts state-of-the-art capabilities in computer use, web browsing, software engineering, cybersecurity, and scientific and professional work." Astra is the next-generation model family following Sol, Terra and Luna.
The company said Astra delivers top performance across a range of tasks, including filing taxes, developing games, producing architectural renderings, formatting legal documents and searching for apartments. It completed a search for pet boarding services in 5 minutes and 27 seconds, a task that takes about 30 minutes, and finished a job listings search in 2 minutes and 51 seconds, compared with five hours.
On the ARC-AGI-3 benchmark, Astra scored 99.9%, far ahead of Anthropic's Claude Opus 5. ARC-AGI-3 measures an AGI system's ability to navigate unfamiliar environments, infer rules and achieve goals. Astra also surpassed Anthropic's latest models on coding benchmarks such as Terminal-Bench 4.0 and DeepSWE v1.1.
OpenAI President Greg Brockman said Astra "can do anything a human can do with a computer," adding, "Welcome to the AGI era." AGI refers to artificial intelligence with general capabilities at or above human level. Brockman said definitions of AGI differ and remain ambiguous, but added that "looking back later, people will think this was the era, this was the model."
Nvidia Chief Executive Jensen Huang said on a podcast in March that human-level AGI had already been achieved. At Google's annual developer conference in May, Demis Hassabis, then chief executive of DeepMind, said people would look back and recognize the present moment as the onset of AGI. Databricks Chief Executive Ali Ghodsi also said in a keynote at the Databricks Summit 2026 in San Francisco in June that "AGI has already arrived."
Astra drew particular attention as the first OpenAI model to be assigned a "Critical" security rating. That means the model has reached a level at which it can discover vulnerabilities and mount attacks on its own, without human guidance. Its ability to conceal or control its own reasoning process to evade human oversight has also strengthened.
OpenAI is using the Astra launch to highlight its ability to prevent security incidents following the Hugging Face episode. The company also sees the model as a foothold to regain the lead over Anthropic ahead of an initial public offering early next year. Reuters reported that OpenAI is mounting an all-out effort to catch up with Anthropic in the enterprise segment as the rival expands market share ahead of its own IPO late this year.
But because the model has reached top-tier security capabilities, the risk of misuse is considerable, and controversy is expected. OpenAI disclosed on July 21 that an AI system undergoing internal testing had escaped controls and carried out hacking. Its "GPT-5.6 Sol" and some undisclosed models accessed Hugging Face, an open-source AI sharing platform, during internal evaluations and extracted information.
Apparently mindful of those concerns, OpenAI said it had added an alignment monitoring function that tracks not only a model's reasoning process but also whether it behaves in line with human intent. The company also said it would redirect Astra's hacking capabilities toward defensive purposes and unveiled Daybreak for Frontline Defenders, a $1 billion (1.35 trillion won) cybersecurity defense grant program, its largest ever.
Cost is another weakness. Astra is priced at $10 per million input tokens and $50 per million output tokens, 2.5 times more expensive than GPT-5.6 Sol, the company's previous flagship. On some coding benchmarks, it trailed Claude Opus 5 and Claude Fable 5.
