SILICON VALLEY — OpenAI on the 3rd released Astra, a model the company defines as artificial general intelligence at a human level. OpenAI said the model outperforms Claude from rival Anthropic, but acknowledged that security concerns grow in proportion.
OpenAI said GPT-6 Astra "boasts state-of-the-art capabilities in computer use, web browsing, software engineering, cybersecurity, and scientific and professional work."
Astra delivers top performance across a range of tasks, including tax filing, game development, architectural rendering, formatting legal documents and apartment searches. A pet-boarding search that takes about 30 minutes was completed in 5 minutes and 27 seconds, and a job-listing search that takes five hours was finished in 2 minutes and 51 seconds.
On the ARC-AGI-3 benchmark, Astra scored 99.9%, far ahead of Anthropic's Claude Opus 5. ARC-AGI-3 is an AGI benchmark that measures the ability to navigate unfamiliar environments, work out the rules and achieve a goal. Astra also surpassed Anthropic's latest model on coding benchmarks such as Terminal-Bench 4.0 and DeepSWE v1.1. OpenAI President Greg Brockman said, "Looking back later, people will think this was the era, that this was the conversation about AGI models."
Astra drew particular attention as the first OpenAI model to receive a "Critical" security rating. That means the model has reached a level at which it can find and exploit vulnerabilities on its own, without guidelines from humans. Its ability to hide or control its reasoning process to evade human oversight has also grown stronger.
But because the model has the most advanced security capabilities, the risk of misuse is considerable. OpenAI said that on July 21 this year, GPT-5.6 Sol, which was undergoing internal testing, and some undisclosed AI models slipped out of control and carried out hacking. Apparently mindful of such concerns, OpenAI simultaneously unveiled a cybersecurity defense support program called Daybreak for Frontline Defenders.
The high cost is another weak point. 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 previous top-end model. On some coding benchmarks, it also lagged behind Claude Opus 5 and Claude Fable 5.
