Breaking

OpenAI Releases GPT-6 Astra for Computer-Use and Agentic Work

September 3, 2026 at 23:59 EDT

Foundation Models
AI Agents
Security

OpenAI released GPT-6 Astra on September 3, positioning it as a frontier model for computer use, coding, research, and end-to-end professional work. The staged rollout begins with a limited group of organizations, with broader ChatGPT and API access scheduled over the following days. OpenAI's launch announcement

SEPTEMBER 3, 2026 / OPENAI

GPT-6 Astra takes computer-use AI into end-to-end work

OpenAI's new model is rolling out for complex computer use, coding, research, and professional workflows, with stronger capability paired with tighter controls.

THE SHIFT IN NUMBERS

72.6%

OSWorld 2.0 score

47%

Less simulated task time

1.05M

Context window tokens

$10 / $50

Input / output per 1M tokens

Astra cuts simulated computer-task time

OSWorld 2.0 latency simulations show higher performance in roughly 40 minutes, versus approximately 75 minutes for GPT-5.6 Sol.

40 min

GPT-6 Astra

75 min

GPT-5.6 Sol

Each block represents 5 minutes of simulated task time.

Why teams are paying attention

REPORTED UPSIDE

Early users report stronger everyday work, coding, legal drafting, browser testing, and fewer production iterations.

CRITICAL CAVEAT

An independent snapshot found a 9.7-point gap between general intelligence and agentic scores, while safeguards can pause legitimate tasks.

Autonomy raises the control burden

Astra's computer-use gains arrive alongside a Critical cybersecurity designation and stronger monitoring requirements.

100.0%

ExploitBench score

2 zero-days

Found during evaluation

0% vs 48%

Unauthorized-scope rate versus Sol without production safeguards

Continue reading

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