OpenAI starts rolling out its next-generation GPT-6 Astra model

OpenAI has started rolling out GPT-6 Astra, its newest large language model, and it’s not being framed as just another incremental update. The company says Astra posts state-of-the-art results in coding, browsing and computer use, the awkward-sounding category for models that have to work through software interfaces like a person would.

The headline numbers are hard to ignore. OpenAI says Astra scored 98% on FrontierMath Tier 4, a benchmark built around 50 math problems that can take human mathematicians weeks. It also hit 99.9% on ARC-AGI-3, which tests how well a model can learn new tasks, and a perfect score on ExploitBench, which measures whether a model can find and use software vulnerabilities.

That last one is exactly why the rollout wasn’t as straightforward as a normal product launch. OpenAI disclosed on Tuesday that Astra counts as a “critical” risk under its internal safety framework, meaning it can hack “many well-protected systems” without human help. The company says that is why release had to be pushed back for several weeks while engineers built guardrails around the model’s hacking abilities.

OpenAI is also pitching Astra as a systems-level improvement, not just a smarter chatbot. The company says it uses a new data management approach that keeps compressed prompt information searchable instead of throwing it away when the context window fills up. The idea is simple: preserve more of the trail, and the model has more to work with later.

For now, access is limited. Astra is available through Daybreak, a program for cybersecurity research, and OpenAI says it will broaden access in the coming days through ChatGPT, Codex and its API. Daybreak participants will get $1 billion in credits, which cover usage, training and customer support, with some of the training done alongside MS-ISAC, the nonprofit that helps public sector organizations block cyberattacks.