I read OpenAI's GPT-6 Astra launch page the way I read every announcement of this kind: first I strip out the adjectives, then I look at the numbers, then I look for the sentences that admit something. The sentences that admit something are almost always the most interesting, and here there are two of them that deserve more attention than all the benchmarks combined.
What OpenAI says
The model is presented as "the world's most intelligent and aligned model". The comparison is almost always with its predecessor, GPT-5.6 Sol, and some tables also include the competing models from Anthropic and Google.
OpenAI leads with three numbers: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. The word they use is "saturate", meaning the model has reached the point where the test no longer measures anything. On the mathematics side there is also a concrete result: an improved bound on gaps between prime numbers (from 240 to 186) and the correction of a term that had stood unchanged for over eighty years.
For anyone working in a business, the part that matters is computer use. On OSWorld 2.0, the test where the model has to operate an operating system the way a person would, Astra reaches 72.6% in about 40 minutes, against 65.7% in about 75 minutes for the previous model. The examples OpenAI lists are deliberately mundane: filling in online forms, updating a CRM, organising a calendar, researching a topic and writing a summary, building a website and checking it. On AutomationBench, which measures exactly this kind of office workflow, the jump is from 18.1% to 41.4%. On BrowseComp, complex online research, the model scores 91.5%.
On code the numbers rise in a similar way (Terminal-Bench 4.0 goes from 37.3% to 57.9%), and the part I care about most is a line from Jane Street, one of the partners quoted: the model "communicates in a way that is easier for developers to follow and produces code that requires fewer iterations". Fewer iterations means less of my time spent re-reading.
Then there is cybersecurity, which OpenAI treats in a different tone from the rest of the page. Astra is the first model the company classifies at the "Critical" level of its own preparedness framework: during testing it found and used two previously unknown zero-day vulnerabilities, and external experts verified that, without production safeguards, it can execute arbitrary code in hardened browsers. This is why the version reaching users blocks advanced offensive activity and only allows code review and patching. Broader access for defensive use is planned over the coming weeks through a programme called Daybreak.
On availability: the rollout starts with a limited group of organisations and over the following days reaches all Plus, Pro, Business and Enterprise users, as well as the API and Amazon Bedrock. For Enterprise customers access is off by default and has to be enabled manually. API pricing is 10 dollars per million input tokens and 50 per million output tokens, with a fast mode that costs twice as much.
The two sentences that matter
The first is in the alignment section. OpenAI writes that Astra's written reasoning is "harder to monitor" than Sol's, based on tests in which the model was explicitly asked to evade monitoring. It adds that the model "still appears to struggle to conceal the reasoning needed for complex tasks", but that the decline is being taken seriously. It is a rare sentence to find on a launch page, and it should be read for what it is: the most capable model is also the one whose inner workings are least legible.
The second is in the section on safeguards. The additional safety checks "can sometimes slow, pause, or stop legitimate work". Translated for anyone who will use it in an office: expect interruptions, including on ordinary tasks, because the classifier system watching the model errs on the side of caution. OpenAI says it is working on this, and I have no reason to doubt it, but in the meantime it is a real operating cost that no table measures.
What changes for anyone running a business
The point I see discussed least is that the announcement no longer describes an assistant that answers. It describes a model that opens the browser, fills in the form, updates the CRM and runs quality checks on the website. The OSWorld test is built on exactly that. And the halved time is not a detail: a task that used to require 75 minutes of automated execution now takes 40, and a 40-minute task can be handed over without much thought, while a 75-minute one has to be scheduled.
This brings me back to something I have been saying for years, and which this announcement makes more urgent, not less. A model that carries out tasks on a computer amplifies whatever it finds. If the CRM has fields filled in three different ways, if the calendar has events with no logic behind them, if the company templates exist only in the head of whoever uses them, the model will work at double speed inside that disorder. There is a line on OpenAI's page that says this without meaning to: Astra produces presentations, documents and spreadsheets "following company templates". The premise is that the templates exist.
In my own week this turns into a concrete question. Of the tasks I do every Monday morning to realign my two brands, which ones have a structure clear enough to be handed to a model that works on its own for 40 minutes? The honest answer, today, is: fewer than half. Not because the model isn't capable, but because the rest still lives in decisions I make on the spot, without ever having written them down anywhere.
One last observation on price. Fifty dollars per million output tokens is not a figure for indiscriminate daily use, and fast mode doubles it. The model makes sense on long, bounded tasks with a verifiable result at the end. For most of what a professional does, a smaller model is still enough, and choosing which model to use for which kind of work becomes a skill in itself.
In practice
If you have a Plus or Business subscription, the model will appear in your list over the next few days. Before trying it on real work, take a workflow you do often and that has a clear beginning and end, write it out step by step, and give it that. The most useful test is not seeing how intelligent it is. It is seeing how much of your work is already structured enough to be handed over.
Which task in your week would you hand first to a model that works alone for 40 minutes? And which one could you not even explain to it?
Alessandra
Source: OpenAI, "Introducing GPT-6 Astra", September 2026. Figures and quotations are taken from the official announcement page; comparisons refer to the previous model, GPT-5.6 Sol.
