Imagine telling your AI "fill out my tax return" – and it actually tries: clicking through forms, entering numbers, even checking the result afterward. That's the promise behind GPT-6 Astra, OpenAI's new flagship model, rolling out gradually since September 3, 2026.
What makes Astra different
Astra isn't just a chatbot update — it's an agent model (AI that carries out several steps on its own instead of just replying). According to OpenAI, it can operate a computer independently: filling out forms, organizing calendars, updating CRM entries (customer databases), building websites, and even running checks on them afterward. It also ships with a 1.05-million-token context window (roughly: several thick novels at once) and up to 128,000 tokens of output. OpenAI itself calls Astra "the world's most intelligent and aligned model" — big words, as usual at a flagship launch.
Free? Not a chance
If you were hoping to try this in the ChatGPT free tier: no luck. Astra is rolling out first to Plus, Pro, Business, and Enterprise subscriptions, plus the API — none of it free at the current stage. Through the API, usage runs in the low double-digit dollars per million tokens, with output costing noticeably more than input. That's nothing unusual: no major provider gives away its best model for nothing — Claude and Gemini keep their top versions behind a paywall too. If you just want to poke around occasionally, GPT-5.6 or a provider with a more generous free tier is still the better bet.
The uncomfortable part: Astra can hack, too
More interesting than the marketing numbers is a footnote in the safety report: Astra is the first model OpenAI itself classifies as "Critical" for cybersecurity. In plain terms: with the right tools, it can independently find previously unknown vulnerabilities in well-protected systems — and ways to exploit them — without a human guiding every step. Because of that, this part of its capability stays restricted for now: only selected testers get full access, with a broader, defensively focused rollout planned through a separate program called "Daybreak Blue."
That sounds more dramatic than it is in everyday use. Most users will never touch this feature — it's built for security research, not for your regular ChatGPT conversation. Still, it shows where things are headed: the more capable an agent gets at "real" action, the bigger its potential for misuse also becomes. Worth keeping in mind, without spiraling into panic.
Who is this for?
Short version: not really for beginners right now. If you already pay for Plus or Pro, Astra ships to you automatically, and you can try the agent features — say, for building a small website or organizing your schedule. Developers working through the API benefit from the large context window for bigger codebases. If you only chat with AI occasionally, you probably won't notice much difference from GPT-5.6 in daily use — and you'll pay extra for it.
The more interesting question isn't "can I use this now" but "do I actually want an AI operating my computer while I get coffee." The answer differs for everyone. Still worth trying once the wider rollout lands — just not for free.