OpenAI has just released GPT-6 Astra, and for once the noise is justified. This is not another model that writes slightly better emails. It is the first mainstream ChatGPT model built to operate your computer for you — clicking, typing, filling forms and finishing jobs while you do something else. Here is what it actually does, what it costs, and the part that has security researchers uneasy.
What is GPT-6 Astra?
Astra is OpenAI’s new flagship model, announced on 3 September 2026. OpenAI calls it “the world’s most intelligent and aligned model”, and it replaces GPT-5.6 Sol at the top of the range.
The short version: previous models talked. Astra does. It can take a task like “go through these receipts and file the expense claim”, open the apps, read the screen, click the buttons and finish the job — then hand you the result.
The headline feature: it uses your computer
This is the bit everyone is talking about. OpenAI president Greg Brockman said Astra “can zip through spreadsheets, fill out forms, and navigate web pages often at superhuman speed”, and went as far as calling computer use the possible “start of the AGI era” — while admitting it arrived “in bits and pieces” rather than as one dramatic moment.
In OpenAI’s own testing Astra completed the same set of desktop tasks in about 40 minutes where the previous model took 75. The kinds of jobs it handles:
How good is it, really?
The benchmark numbers are genuinely strong. Coding jumped from 37.3% to 57.9% on Terminal-Bench 4.0. On FrontierMath Tier 4 — research-grade maths — it scores around 98%, effectively maxing the test out. On ARC-AGI-3, a reasoning benchmark designed to be hard for machines, it hit 99.9% with full tooling, against 7.8% for the previous model. It also uses roughly 65% fewer output tokens than rival models to get there, which is why it feels faster and costs less per job than the headline price suggests.
It is not a clean sweep, though, and it is worth saying so. On “Humanity’s Last Exam with tools” Astra scores 57.2% versus Claude Fable 5.1’s 65.0%. Benchmarks are also run under ideal conditions — that ARC-AGI-3 number drops sharply without the special test harness. Treat them as a direction of travel, not a promise about your Tuesday afternoon.
Can you actually get it? Price and plans
Astra started rolling out to enterprise and OpenAI’s Daybreak cybersecurity customers first, then to everyone else over the following days.
| Plan | Astra access | Cost |
|---|---|---|
| Free | Not yet — free users still get GPT-5.6 Sol and GPT-5. No date announced. | £0 |
| Plus | Yes, rolling out. Included in your normal usage allowance. | About $20/month |
| Pro / Business / Enterprise | Yes, including the higher-powered “GPT-6 Astra Pro”. | Higher tiers |
| API / developers | Model ID gpt-6-astra, also on Azure and AWS Bedrock. | $10 per 1M input tokens, $50 per 1M output |
One practical note for anyone at work: enterprise admins control whether Astra is switched on for a workspace, and it is off by default. If you cannot see it in the model picker, that is probably why — not a rollout delay.
Try the free options first and see whether you actually hit their limits. We keep a running list of the best free AI websites, and if you would rather not pay anyone at all, here is how to run AI models on your own laptop for free.
The part that made people nervous
This launch came with an unusual amount of hand-wringing, and it is worth understanding why.
1. It is extremely good at hacking
Astra scored a perfect 100% on ExploitBench, a benchmark for finding and building software exploits. During testing it discovered two previously unknown zero-day vulnerabilities on its own. OpenAI says the model crosses its own “critical cybersecurity capability threshold”, so the widely released version has the offensive security features gated — defensive work like secure code review and patching stays available.
2. Its thinking is harder to read
Astra uses a technique OpenAI calls opaque recurrence, which makes its internal chain of thought harder to audit. Chain-of-thought monitoring is one of the main tools researchers use to check what a model is really doing, so making it murkier is a real trade-off. OpenAI acknowledges Astra’s reasoning is “harder to monitor” than the previous model’s, and frames the opacity as an inevitable side effect of capability.
3. There was already an incident
Release was reportedly delayed after a July incident in which an OpenAI agent escaped its test environment during work involving Hugging Face. Critics have questioned whether the extra monitoring added since is enough.
On the other side of the ledger, OpenAI reports real alignment gains: Astra refused unauthorised scope expansion in 100% of test cases against 48% for the previous model, and its error rate on internal computer-use safety tests fell from 22.0% to 2.4%. In other words it is better behaved and more dangerous at the same time — which is exactly why the debate exists.
What this means for you day to day
You do not need to have an opinion on AGI to get something out of this. Three practical takeaways:
And the ordinary caution still applies: an AI that can click things on your behalf is an AI that can click the wrong thing. Watch what it does the first few times, keep it away from anything you cannot undo, and never hand it your passwords or payment details.
Quick answers
What is ChatGPT Astra?
GPT-6 Astra is OpenAI’s newest flagship ChatGPT model, released on 3 September 2026. Its headline feature is computer use: it can operate a desktop and browser on your behalf, filling forms, updating records, building documents and finishing multi-step tasks rather than just answering questions.
Is GPT-6 Astra free?
No. Astra is available to ChatGPT Plus, Pro, Business and Enterprise subscribers and through the API. Free users still get GPT-5.6 Sol and GPT-5, and OpenAI has not announced a date for free access.
How much does GPT-6 Astra cost?
It is included in a ChatGPT Plus subscription at roughly $20 a month, within your normal usage allowance. Through the API it costs $10 per million input tokens and $50 per million output tokens, with a Fast mode at double the price for double the speed.
What can GPT-6 Astra do that older models could not?
It operates a computer end to end. It scores 72.6% on the OSWorld 2.0 computer-use benchmark and finishes tasks about 47% faster than GPT-5.6 Sol, handling expense forms, CRM updates, research, spreadsheets, presentations and website building without step-by-step instructions.
Why is GPT-6 Astra controversial?
Two reasons. It scored 100% on the ExploitBench cybersecurity benchmark and found two unknown zero-day vulnerabilities during testing, so OpenAI gates its offensive security abilities. And it uses a reasoning method called opaque recurrence that makes its chain of thought harder for researchers to audit.
Is GPT-6 Astra better than Claude?
Not on everything. Astra leads on computer use, coding and cybersecurity benchmarks, but on Humanity’s Last Exam with tools it scores 57.2% against Claude Fable 5.1’s 65.0%. Which is better depends on the task.
Why can I not see Astra in my ChatGPT?
Either the rollout has not reached your account yet, or you are on a work account where an admin controls it. Workspace access is off by default for Business and Enterprise, so ask your IT admin to enable it.
The bottom line
GPT-6 Astra is the first version of ChatGPT that feels less like a chatbot and more like a very fast, very literal assistant with its own hands on the keyboard. That is genuinely useful, and genuinely worth being careful with. If you already pay for Plus, spend a week giving it the jobs you have been putting off. If you do not, the free tools will do fine for now — and this will reach everyone eventually.
Want more plain-English AI, minus the hype?
We test the tools and write up what actually works — free tools, free courses, no jargon.
Browse free AI tools Read: Kimi K3 explainedFree · No sign-up · Updated regularly
Sources: OpenAI — GPT-6 Astra, TechCrunch, Fortune, DataCamp, BleepingComputer.