Breaking

OpenAI's GPT-6 Astra Runs Slower but Leaner in Repeat Test Days After Launch

September 12, 2026 at 11:45 EDT

Foundation Models
Security
AI Agents

A side-by-side test of GPT-6 Astra on an identical task found OpenAI's newest flagship model taking longer to finish the job six days after its launch while using noticeably fewer tokens to complete it. The gap adds a concrete data point to a broader debate over whether frontier AI systems change in meaningful ways within days of shipping.

Sep 11, 2026 · OpenAI

GPT-6 Astra: Slower, Leaner, One Week On

A repeat test of the same task found OpenAI's new flagship model taking longer to finish six days after launch while using noticeably fewer tokens to do it.

Same prompt, tested six days apart

Time To Complete

26 min

Sep 5

30 min

Sep 11

Tokens Consumed

10M

Sep 5

Sep 11

Same task, run six days apart: the later pass took longer but spent fewer tokens.

Read as efficiency

Astra's recurrent-depth design was built to do more with fewer tokens, and OpenAI says its computation depth stays close to prior models.

Read as a warning sign

Usage caps for heavy users were cut by up to 4x days after launch, and OpenAI's own tests show reasoning traces have grown harder to monitor.

Why It Matters

Astra is OpenAI's first 'Critical'-tier model

• Rated able to find and exploit unknown security flaws with minimal human help

• Chain-of-thought reasoning has become shorter and harder to monitor

• Full capabilities remain gated to vetted testers under restricted rollout

Continue reading

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