OpenAI shipped GPT-6 Astra on 3 September. Greg Brockman called it a generational leap and said it might eventually be read as the arrival of AGI. Sam Altman said it was a new capability level and that it had already changed how he works.
Then most people went to open ChatGPT and found the same model picker they had the day before.
That gap is the thing I want to talk about, because we have now watched this exact movie four or five times and it is starting to get on my nerves.
What actually shipped
Astra is GPT-6. The naming argument that ran for months is over, and the API model ID is gpt-6-astra. It has a one million token context window. Pricing is $10 per million input tokens and $50 per million output, with cached input at $1 and batch at half rate. That is roughly 2.5 times the per-token cost of GPT-5.6 Sol.
OpenAI positioned it around computer use and long-running work rather than conversation. Browsing, software engineering, multi-step professional tasks, producing actual documents and spreadsheets instead of telling you how to produce them. Computer use is close to twice as fast as before, and the harness optimisations that came with it apparently speed up Sol by about 60 per cent as well. Codex also got a new way to preserve and retrieve context when the window fills, instead of the usual summarise-and-hope compaction.
The benchmark table was the loud part. 98 per cent on FrontierMath Tier 4. 99.9 per cent on ARC-AGI-3. 100 per cent on ExploitBench. OpenAI also published a Lean-formalised proof improving a bound on prime gaps that had held for more than 80 years, which is a genuinely lovely thing to ship on launch day and got about a tenth of the attention it deserved.
The part that annoys me
Here is the sequence. Announcement day arrives with AGI framing, a benchmark table, and partner demos. Access goes to a small set of vetted organisations. The release notes the next day still say the model is not generally available. Paid customers, including Pro subscribers paying $200 a month, sit there refreshing.
Altman ended up apologising and saying Pro would go first in the wider rollout. Tibo Sottiaux, who runs Codex and ChatGPT, offered a banked limit reset for every day a paid account went without access. Plus users who did get something found it only inside ChatGPT Work and Codex rather than the normal model picker, which produced exactly the confusion you would expect.
I want to be fair about the reason, because it is a real one. OpenAI classified Astra at the Critical threshold for cybersecurity under its own Preparedness Framework. That means a suitably equipped version can find unknown vulnerabilities and build exploits against well-defended systems without a human walking it through each step. The company says it found two zero-days during testing. Staging that carefully is defensible. I would stage it too.
But safety staging and clear communication are not in conflict. "Limited security preview, wider access in two weeks" is a sentence anyone can understand. Announcing a model as a public event, framing it as the start of the AGI era, publishing a benchmark page and then taking it down hours later, and running access like a private members club is a different thing. It gets you the cultural moment without the operational cost of actually serving everyone.
The other cost is smaller but real. When labs do this repeatedly, the people who have to build on top of the model learn to discount every launch by default. We now treat capability, price and availability as three separate variables that have nothing to do with each other, and we write code that never depends on a single model name. That is good engineering hygiene, but the labs taught it to us the hard way.
The benchmark asterisks are worth reading
If you only read the headline numbers you get one story. Independent testing gives you another.
Artificial Analysis measured Astra at 61.2 on its broad Intelligence Index against 60.9 for GPT-5.6 Sol. That is parity, at 2.5 times the token price. Coding moved more meaningfully, 67.0 against 65.1 on the Coding Agent Index, but it is not the leap the marketing implies. Some observers put it behind an open-weight competitor on the same index.
The ARC-AGI-3 result has a footnote too. The 99.9 per cent figure came from a Provider Adapter harness that keeps private reasoning state and compacts long conversations. ARC Prize reported 62.7 per cent using its standard harness. Both numbers are real, they just measure different setups, and only one of them was on the slide.
None of this makes Astra a bad model. It makes the launch claim and the measured reality two different objects. Those numbers came from independent evaluators, not from cynics on X, and if you are budgeting for a migration you should be reading them before you read the announcement post.
And yet: the demos look fantastic
I am going to drop the scepticism for a paragraph, because it would be dishonest to pretend otherwise.
The computer use demos are the most convincing thing I have seen from any lab this year. Not a model telling you which buttons to press, but a model pressing them, in the background, while you work on something else. Ethan Mollick had early access and built a simulated Library of Alexandria that people are pointing to as evidence of real multi-day autonomous work. The design and output quality in the previews, the documents and slides and spreadsheets it produces without hand-holding, are noticeably better than what we are used to cleaning up afterwards.
We are genuinely excited to get our hands on it. Long-horizon agentic work is the exact shape of problem we keep running into with clients: the process that takes a person four hours and eleven browser tabs and three systems that do not talk to each other. If Astra can hold context across a job that long without falling over, that changes what we can quote for.
So yes, we will be testing it the moment it lands properly in our accounts. We will also be measuring it ourselves rather than taking the benchmark table on faith.
What we will actually be checking
Cost per completed task, not cost per token. A 2.5x token price is irrelevant if the model finishes the job in one pass instead of four. It is very relevant if it does not. OpenAI has argued the former, but has not published enough task-level detail for anyone to verify it, so we will work it out on our own workloads.
Reliability over long runs. Every model looks good for ten minutes. The question is what happens at hour three of a document processing job at two in the morning.
The safety interrupts. OpenAI has acknowledged that the monitoring built around Astra's cyber capabilities can pause or stop legitimate work. For a consultancy doing integration and infrastructure, that is not an abstract concern. We need to know what trips it.
Whether it actually beats what we already run. Parity on a general intelligence index at 2.5 times the price is not automatically an upgrade. For a lot of the automation we build for accounting firms and Australian SMBs, the cheaper model was already good enough, and the sensible architecture is an endpoint you can repoint rather than a rewrite.
Frequently asked questions
Is Astra the same thing as GPT-6?
Yes. OpenAI's official name is GPT-6 Astra and the API model ID is gpt-6-astra. The two names refer to one model.
When was GPT-6 Astra released?
3 September 2026, first to a limited set of organisations in OpenAI's Daybreak programme, with paid ChatGPT tiers, the OpenAI API and AWS following over the days after.
How much does GPT-6 Astra cost?
$10 per million input tokens and $50 per million output on the API. Cached input is $1, batch and flex processing run at half price, and Fast mode costs double the standard rate.
Why was the rollout staged?
OpenAI classified Astra at the Critical cybersecurity threshold under its Preparedness Framework, meaning it can find unknown vulnerabilities and develop exploits without step-by-step human direction. Broader offensive security capability is gated behind the vetted Daybreak programme.
Should we migrate our automations to Astra?
Not on the announcement alone. Measure cost per completed task on your own workloads first, and keep your integration behind an interface you can repoint if the numbers do not hold up.
The short version
Astra looks like a real step forward on agentic and computer-use work, a marginal one on general reasoning, and a significant one on price. The launch was handled as a marketing event first and a product release second, which is a pattern the whole industry has settled into and which I would like someone to break.
If you are running a business and wondering whether any of this changes what you should be doing on Monday: not immediately. It will, in a few months, in ways that are easier to plan for than to react to. That is usually the useful window.
OrionX Technologies builds AI automation, custom software and cloud infrastructure for accounting firms and Australian SMBs. If you want to talk through what agentic AI actually changes for your operations, get in touch.
