Breaking

OpenAI Launches GPT-6 Astra With Agentic Computer Use and Guarded Cyber Access

September 3, 2026 at 15:17 EDT

Foundation Models
AI Agents
Security

OpenAI has launched GPT-6 Astra, a new flagship model aimed at operating computers and browsers, writing software, solving advanced mathematics, and handling longer multi-application workflows. The release also makes a safety milestone part of the product story: OpenAI says Astra is the first model it has designated as meeting its Critical cybersecurity capability threshold.

September 3, 2026 · OpenAI

GPT-6 Astra moves AI from answers to action

OpenAI's new flagship model is built to navigate software, complete multi-step work, and operate with unusually advanced cybersecurity capability.

The release at a glance

100%

ExploitBench score

91.5%

Cyber jailbreak refusal

72.6%

OSWorld 2.0 offline

$50

Per 1M output tokens

Refusal is part of the capability story

OpenAI reports a higher refusal rate for disallowed cyber assistance in its internal evaluation.

GPT-6 Astra
91.5%

GPT-5.6 Sol
59%

Higher is better for refusal rate.

What users gain

Astra can navigate browsers, spreadsheets, desktop tools, and longer workflows. The strongest early use cases pair the model with an environment where it can see, act, test, and revise.

What remains hard

More autonomy raises the cost of mistakes. Monitoring can slow, pause, or stop legitimate work, while shorter reasoning traces can make behavior harder to inspect.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.