OpenAI Releases GPT-6 Astra, Claims AGI Has Arrived

OpenAI shipped GPT-6 Astra on September 3, 2026, and Greg Brockman went on stage to say what the company had been carefully avoiding for years: “Welcome to the AGI era.”

He walked it back slightly. AGI is “a much more gray, fuzzy thing” than a single defining moment, Brockman acknowledged. But the framing was deliberate. This is the model OpenAI is treating as the threshold.

What it actually does

Astra is built around computer use. It navigates software the way a person would: clicking through browsers, filling out forms, updating CRM records, editing spreadsheets, running engineering tools like KiCad and FreeCAD. It doesn’t just tell you how to do things. It does them.

On OpenAI’s OSWorld 2.0 benchmark, Astra completes tasks in about 40 minutes on average, versus 75 minutes for GPT-5.6 Sol. That’s a 47% speed improvement on the same set of real desktop tasks.

The benchmark scores are striking:

ARC-AGI-3: 98.6%
FrontierMath Tier 4: 97.6%
DeepSWE v1.1 (bug fixing): 74.1%
ExploitBench: 100%
OSWorld 2.0: 72.6%
Terminal-Bench 4.0: 57.9%

The 100% on ExploitBench is the number that’s making people nervous. Astra can find and exploit security vulnerabilities autonomously, which is why it’s the first OpenAI model to hit the “Critical” threshold under the company’s Preparedness Framework.

The cybersecurity problem

OpenAI isn’t handing this out freely. Initial access goes to enterprise customers enrolled in a gated program called Daybreak Blue. Administrators have to manually enable the model; it’s off by default. Paid ChatGPT plans and API access are following in the days after launch.

The company calls Astra’s written reasoning “harder to monitor” than GPT-5.6 Sol’s. They attribute that to faster problem-solving, but acknowledge monitorability is still an open research problem. That’s a notable admission for a model they’re calling a potential AGI milestone.

What’s changed for developers

The context window is 1,050,000 tokens with 128K maximum output. Knowledge cutoff is April 30, 2026.

Codex users get a meaningful upgrade: instead of compressing long sessions into summaries, Astra maintains searchable notes across context windows. Earlier requirements, failed fix details, and test results stay retrievable rather than getting flattened into a recap. This addresses one of the more frustrating limitations in long coding sessions.

API pricing is $60 per million tokens standard, $120 for fast mode. OpenAI is framing the value proposition around cost per completed task rather than token counts, which is a reasonable way to think about a model this capable of autonomous work.

GitHub Copilot already has Astra available as of September 4, for all plan tiers.

The AGI question

OpenAI’s GDPval benchmark, designed explicitly to measure economically meaningful real-world work, was conspicuously absent from launch materials. That’s the benchmark that would make the AGI claim verifiable in concrete terms. Its absence doesn’t undermine Astra’s capabilities, but it does leave the “AGI era” framing without a precise anchor.

What’s not in question: Astra is a significant leap on both benchmarks and practical tasks. Developers will notice the difference. Whether that constitutes AGI depends on what definition you’re working from, which is precisely Brockman’s point.

Sources: VentureBeat, DataNorth AI, 9to5Mac