Breaking

OpenAI Launches GPT-6 Astra With Recurrent Depth and a Monitorability Trade-Off

September 5, 2026 at 11:19 EDT

Foundation Models
AI Agents
Security

OpenAI has launched GPT-6 Astra, a frontier model aimed at long, multi-step work in coding, computer use, research and cybersecurity. The September 2026 release pairs higher benchmark performance and shorter task times with an admission that Astra's written reasoning is harder to monitor than GPT-5.6 Sol's.

SEPTEMBER 2026 · OPENAI

More capability. Less visible reasoning.

GPT-6 Astra raises coding and computer-use performance while reducing the amount of reasoning that conventional monitors can inspect.

72.6%

OSWorld 2.0 score

40 min

Simulated task time

57.9%

Terminal-Bench 4.0

$10 / $50

Input / output per 1M tokens

THE OVERSIGHT GAP

Astra can shape its reasoning trail

CoT-Control results in the 750-to-1,250-token range. Taller columns show more successful control of the written chain of thought.

60.9%
GPT-6 Astra

16.1%
GPT-5.6 Sol

Capability upside

Higher scores in computer use and coding, plus roughly 47% less simulated OSWorld task time.

Oversight cost

Shorter, less informative chains of thought can weaken CoT-only monitoring, while safeguards may interrupt legitimate work.

Continue reading

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