Breaking

OpenAI’s Astra May Match a GPT-4-Scale Leap as Recurrent-Depth Reasoning Spreads

September 2, 2026 at 10:34 EDT

Foundation Models
Security

OpenAI’s forthcoming Astra model could deliver a performance jump comparable to the leap from GPT-3.5 to GPT-4, according to researchers tracking its coding and computer-use results, and the gain is expected to rest in part on new reasoning methods now spreading across major labs. The company has said it plans to make Astra available soon, while restricting the model’s strongest cybersecurity tools.

September 2026 · OpenAI Astra

A GPT-4-scale leap, with Critical cyber risk

OpenAI’s forthcoming Astra may match a GPT-4-era jump in coding and computer use, using reported recurrent-depth reasoning—while becoming the first model the lab rates Critical for cybersecurity.

Cyber refusal rates (internal tests, no full production safeguards)

59%

GPT-5.6 Sol

91.5%

Astra

Stacked blocks are proportional (~6 vs ~9) to the measured refusal rates.

100%

ExploitBench (known vulns)

Zero-days in internal eval

Open math problems (~$2,000)

~2×

Complexity vs GPT-4 (Pachocki)

How reported recurrent depth works

1 Input text

→

2 Same layers, looped

→

3 Next token + readable CoT (throttled)

Capability case

Researchers say coding and computer-use gains could resemble the 2023 GPT-4 jump; a smaller model can act like a larger one on math and code.

Oversight case

Safety researchers worry hidden-state loops make chain-of-thought monitors less reliable if later models drop the throttle.

Access split

Release soon; strongest cyber tools stay gated

Default product: extra monitors that can pause even non-cyber work. Daybreak Blue testers: expanded cyber capability used in the Critical-threshold evals.

Continue reading

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