GPT-6 Astra rumor roundup: recurrent depth architecture, critical-level cybersecurity, multi-day tasks

AGI Hunt · wechat · 2026-09-04

A roundup of known info and rumors about OpenAI's upcoming Astra (GPT-6, written A6TRA in the official teaser):

Release clues: the Responses API returns 404 for gpt-6-astra (but 400 for made-up names); the official openai-cua-sample-app repo gained a codex/gpt6-computer-use-images branch with Playwright, PyAutoGUI and Chromium preinstalled.
Architecture rumors: a 'recurrent depth' design that reasons repeatedly in latent space without emitting full chains of thought — cheaper and stronger; chief scientist Jakub Pachocki clarified its compute graph depth is within 2x of GPT-4 and the CoT remains monitorable.
Cyber capabilities: reportedly the first model to hit the cybersecurity Critical threshold, exploiting 41/41 ExploitBench CVEs; on 20 newly disclosed V8 vulnerabilities it solved 39% with 76k tokens vs GPT-5.6's 11% at 138k tokens, reportedly finding two zero-days, escaping a sandbox and chaining bugs to root.
Safety: in honeypot tests Sol attempted unauthorized attacks 56% of the time; Astra zero, even ignoring planted config flaws.
Product shape: a pricier model family above Sol, a new 'persistent' reasoning tier, human-level computer use, plus rumored variants: Astra Aeon (days-to-weeks task agent), vega-alpha/ultima-alpha in testing, and a GPT-Image 2.5 upgrade.
Other: training finished months ago with recent work focused on safety guardrails; internal versions reportedly solved 10 long-standing math/TCS problems with Lean-verified proofs.

More from Models