Breaking

OpenAI begins GPT-6 Astra rollout as executives say the AGI era has arrived

September 3, 2026 at 22:02 EDT

Foundation Models
AI Agents
Security

OpenAI on Thursday began rolling out GPT-6 Astra, a new flagship it calls its most intelligent and aligned model, as president Greg Brockman told reporters the launch is a “generational leap” that brings the frontier into the “AGI era.”

September 3, 2026 · OpenAI

GPT-6 Astra enters the AGI argument

OpenAI began a limited rollout of GPT-6 Astra, its new flagship for computer use, coding, and professional work. Executives say the AGI era has arrived; independent evaluators say a saturated test is not proof.

99.9%

ARC-AGI-3 with provider adapter harness

100%

ExploitBench vs 78.5% for GPT-5.6 Sol

1.05M

Token context window

$10 / $50

API price per 1M input / output tokens

Desktop tasks, drawn to scale

OSWorld 2.0 latency simulations: average minutes per task. Column height matches the time.

~75 min

GPT-5.6 Sol
65.7% score

~40 min

GPT-6 Astra
72.6% score

About 47% less time per task — 15 stacked blocks vs 8, one block per 5 minutes.

What testers like

Computer use is treated as a core capability: filling forms, driving KiCad and Blender, finishing long jobs with fewer stalled loops. Codex can keep searchable notes and ask clarifying questions without blocking other work.

The caveats

The 99.9% ARC-AGI-3 figure uses a provider harness; the shared-interface score is 62.7%. Written reasoning is harder to monitor than Sol’s. DeepSWE sits at 74.1% versus 75.4% for Muse Spark 1.3. API cyber checks can hard-stop a task.

Critical cyber tier

First OpenAI model at the Critical cybersecurity threshold

Without production safeguards, Astra found two previously unknown zero-days in evaluation. Standard access refuses exploit discovery; vetted Daybreak defenders get broader tools first.

1 · Daybreak / trusted access → 2 · Plus, Pro, Business, Enterprise → 3 · API, AWS, Azure

Continue reading

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