Last updated: 2026-09-17 — GPT-6 Astra shipped on September 3, 2026. This article was written before launch, while the release was on hold; the lead, tenses, developer advice and FAQ have been updated to reflect how the delay ended. Current specs: What is GPT-6 Astra.
Quick Answer
GPT-6 Astra shipped on September 3, 2026, after a delay that traced to one thing: its cybersecurity capabilities rated "Critical" — exceeding OpenAI's internal safety threshold for offensive network/exploit capability. On August 7, 2026, release was slowed into isolated testing; the model cleared that process and shipped on September 3. This wasn't "the model isn't ready" — it was "the capability is strong enough that the safety team won't release it unguarded."
What "Critical" Means
OpenAI scores frontier models on safety dimensions, with cyber as one axis. Critical is the top tier, roughly meaning the model's demonstrated capabilities approach or reach "exploitable in real-world attacks" — autonomously finding vulnerabilities, constructing exploit chains, assisting penetration, and so on.
- It does not mean the model is malicious.
- It means the capability floor is high enough to require control, because the same ability is dangerous in a malicious user's hands.
- The standard response is isolated testing + limited release + added guardrails, not cancellation.
For a refresher on the model behind the headline, see What is GPT-6 Astra — the review was about release gating, not about the model itself.
Why This Slowed Release
- Evaluation takes time — full red-team assessment across safety/cyber/bio/chem dimensions runs weeks.
- Critical triggers re-testing — a flagged dimension forces a second, stricter pass plus guardrail validation.
- New policy process — Astra was the first frontier model evaluated under the Trump administration's AI reporting framework, a new pipeline that was still being ironed out.
So "paper Aug 1 → Critical Aug 7 → delayed" was really "new capability meets new regulation."
Historical Precedent
This wasn't unprecedented. GPT-5 went through similar safety-review delays, and OpenAI's pattern is "evaluate → add guardrails → release conditionally." Before launch, the most likely reading of Astra's delay was "adding guardrails, preparing conditional release," not "project in trouble." The market's ~72% September odds reflected that read of "delay, not cancellation" — and it proved right: Astra shipped on September 3.
What Developers Should Expect
- Don't bet on a date — the pre-launch advice was to prep the access layer rather than tie plans to a specific release day; Astra ended up shipping on September 3, and the advice holds for the next launch.
- Check the shipped numbers, not the showcase — before release we expected some high-risk capabilities to be restricted or guardrailed, so that the shipped model might differ from the paper's showcase results. The shipped model's cyber scores (ExploitBench 100%, ExploitGym 42.4%) are vendor-reported — see the published benchmarks rather than relying on pre-launch material.
- Tighter risk control — new flagships launch with throttled quota and stricter risk control; use your own key and stable egress.
FAQ
Q: Does "Critical" mean Astra is dangerous? It means "capability reached a level requiring management," which is different from "malicious intent." OpenAI's response was additional review and isolated testing before release, not destruction.
Q: Did this cancel the release? No. Astra cleared isolated testing and shipped on September 3, 2026. Safety reviews end in "delay + release," not cancellation — GPT-5's history pointed the same way.
Q: Should developers avoid Astra because of this?
No. The model is available now (API model ID gpt-6-astra, plus ChatGPT Plus/Business/Pro/Enterprise). Just set rational expectations: risk control around a new flagship can be tight early on. Prep the access layer — How to use GPT-6 Astra covers the setup — and use your own key.
Summary
Astra's delay was "strong capability + new regulatory process," a delay rather than a cancellation: it ended with the September 3 release. The developer move wasn't watching the controversy — it was preparing an OpenAI-compatible access layer, which now only needs the model name gpt-6-astra. Sign up for Moyi API and start using it today.
Get Started
Moyi API — your OpenAI-compatible entry, with GPT-6 Astra available now.