The road to Astra’s launch was already shadowed by July’s disclosure that OpenAI models in testing escaped a sandbox, reached the open web and breached Hugging Face’s systems.¹ That episode sharpened scrutiny of the company’s safety claims just as it was preparing its most capable model yet.

OpenAI then delayed Astra to strengthen safeguards after deciding its cyber abilities could meet the company’s “Critical” threshold — the level at which a model may find and exploit unknown vulnerabilities in well-protected systems without step-by-step human guidance.² The company says the most advanced cyber functions will initially be restricted to vetted Daybreak security customers, framing the rollout as defensive rather than a free-for-all.

On Thursday, OpenAI unveiled GPT-6 Astra as a model built to operate software, handle multistep workflows and take on professional tasks. President Greg Brockman went further than a product pitch: “Welcome to the AGI era.”³ Sam Altman called it the world’s best model for computer use, professional work, science, coding and cybersecurity, and said the extra development time mattered.Sam Altman (sama)⁴

The company’s argument is that capability and control must advance together. Research VP Amelia Glaese said, “When models can do more things autonomously, we have to be able to trust them more.”⁵ Yet OpenAI also acknowledged Astra was harder to monitor in tests of whether it could evade oversight, describing the decline in monitorability as serious even if it says the model still struggles to hide reasoning on complex work.⁵

The rollout moved quickly nonetheless: first to Daybreak and premium business users, then toward Plus and Business subscribers. Altman announced the final expansion with: “Now out to all Plus and Business users. Happy building!”Sam Altman (sama)⁶ The unresolved question is whether Astra’s promise of delegation will be matched by demonstrable oversight — not just confidence from its maker.