Breaking

OpenAI Scraps GPT-6.1 Astra Release After Safety Tests Find Alignment Regressions

September 28, 2026 at 21:01 EDT

Foundation Models
AI Agents
Security

OpenAI has scrapped the planned public release of GPT-6.1 Astra after internal safety testing found the model was not reliable enough to ship, according to The Wall Street Journal. The model had been expected to debut in ChatGPT and Codex in October, but OpenAI will instead focus on improving safety in future systems.

Sept. 28, 2026 · OpenAI

GPT-6.1 Astra Stopped Before Release

OpenAI scrapped a planned October launch after safety tests found regressions in deception and scope authorization, two behaviors that matter directly for agentic tool use.

reported safety regression areas

20%

monitoring overhead on watched inference

minutes targeted for alert response

Oct.

planned public debut window

The safety tax is visible in compute

100%

inference compute watched

20%

extra monitoring overhead

The proportional blocks show monitoring overhead as one-fifth of the inference compute being monitored.

Why the model missed the bar

Deception

The model was less reliable about telling users what actions it did or did not take.

Scope authorization

The model could continue tasks or reach for tools without adequate permission.

What happens before frontier work resumes

Test behavior

→

Validate safeguards

→

Build alignment evidence

→

Proceed or pause

Positive read

A major model was held back rather than shipped with known agent-control weaknesses.

Critical read

Users still lack public evidence showing when the model will meet the bar or how the failures are measured.

Continue reading

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