OpenAI's next flagship — widely expected to ship as GPT-6 — is called Astra. Here's what's actually confirmed versus rumor, and what it means for teams building on AI today.
Confirmed timeline
- Aug 1: OpenAI confirmed Astra as its "next major model" and said it solved 10 long-open math problems.
- Aug 7: Work on Astra was partly paused after evals could not rule out a "Critical" cybersecurity rating.
- Sep 1: OpenAI confirmed Astra is "available soon" and is the first OpenAI model to genuinely hit "Critical" under its Preparedness Framework.
Why "Critical" matters
Under OpenAI's framework, that rating means Astra can independently find and exploit zero-day vulnerabilities in many hardened live systems with no human help. Its predecessor, GPT-5.6 Sol, was only rated "High." To compensate, OpenAI will restrict Astra's most advanced cyber skills to vetted test partners and emphasize blocking malicious use and misalignment.
The magic: recurrent depth
Reports describe a novel "looped transformer" architecture that reasons in latent space rather than readable text — multiple passes improving the answer while hiding the chain-of-thought. Giving it great brain power while drawing concern from researchers who lose visibility into its reasoning.
Is it really GPT-6?
OpenAI has not announced a GPT-6, price, or date. Astra may launch as GPT-6, GPT-7, or as an own model. Treat them as likely the same until OpenAI says otherwise. Prediction markets put release ~60% before Sep 15.
Current models until then
- 🟢 Sol (flagship) — highest rate
- 🟢 Terra (mid) — mid rate
- 🟢 Luna (budget) — lowest rate
The GPT-5.6 family (Jul 9) is still OpenAI's strongest cyber model, scoring 80 on the Coding Agent Index.
Advice for Gulf teams
Build on what ships today behind OpenAI-compatible endpoints so swapping in Astra later stays a config change. Watch cost-per-task, not list price, and put clear data governance around agentic features.