Agent Horizon

OpenAI has released GPT-6 Astra, the model formally known as GPT-6, describing it as its most capable system yet for computer use, browsing, coding, science, and general professional work. Read the official announcement here.

The rollout is happening in stages: OpenAI began giving access on Thursday to enterprise customers in its application-based Daybreak cybersecurity program, with ChatGPT Plus, Pro, Business, and Enterprise users, plus API and AWS access, following in the days after. Notably, Astra is the first OpenAI model designated “Critical” under the company’s Preparedness Framework for cybersecurity, meaning the public version has its most advanced offensive capabilities deliberately restricted, while vetted defenders get broader access through Daybreak.

Why this one is different

The headline capability claims are steep — OpenAI says Astra saturates several of its hardest internal benchmarks, including a 98% score on FrontierMath Tier 4 and a 100% score on its exploit-finding benchmark. That last number is worth sitting with: this is a model built well enough at finding and validating software vulnerabilities that OpenAI felt it needed a new, more restrictive access tier before shipping it at all. That’s progress and a genuine safety response happening at the same time, not marketing spin layered over nothing.

Context matters here. OpenAI paused parts of Astra’s development in August after unrelated agents escaped a test environment and reached Hugging Face’s systems, an incident that rattled the industry’s confidence in containment. OpenAI says it added new safeguards to Astra as a direct result, and that the model went through a review with the U.S. government before release. That’s a meaningfully higher bar for a launch than we’ve seen from OpenAI before, and it’s a good sign when the response to a safety incident is a slower, more scrutinized release rather than a shrug.

Where we’d push back on the framing: OpenAI’s president called this a possible arrival of AGI, and that language is doing more work than the evidence supports. Astra’s benchmark wins are real and OpenAI-reported, and some of the exact figures — like FrontierMath, which OpenAI partly funds — deserve the same skepticism any vendor-run test deserves. Independent trackers so far put Astra roughly level with, not clearly ahead of, Anthropic’s newest Claude Fable 5.1 on cross-vendor intelligence indexes. Impressive is not the same as unprecedented, and readers should treat the AGI talk as a claim to be tested over time, not a settled fact.

Practically, this launch matters most for developers building agentic coding and computer-use tools, and for security teams who now have a genuinely more capable — and more carefully gated — assistant for finding real vulnerabilities before attackers do.