Estimated read time: 8 minutes
OpenAI shipped GPT-6 Astra on September 3, and by the time you read this it is rolling into ChatGPT Plus, Pro, Business and Enterprise accounts. The headline numbers are genuinely impressive. The price is genuinely alarming. And the gap between those two facts is where most small business owners are going to make an expensive mistake this month.
So let us do the unglamorous thing and separate what actually changed for a five-person company from what changed for an AI lab with a nine-figure compute budget.
Table of Contents
What OpenAI actually shipped
GPT-6 Astra is OpenAI’s new flagship, with an API model ID of gpt-6-astra, a context window of 1,050,000 tokens, a maximum output of 128,000 tokens, and a knowledge cutoff of April 30, 2026. It takes text and images as input. Audio and video are not supported at launch, which is worth knowing before you plan a workflow around it.
The rollout is staged. A limited set of organizations got it on day one through a trusted-access program. ChatGPT Plus, Pro, Business and Enterprise follow over the coming days, with Enterprise switched off by default until an admin turns it on. Broader API access is listed as coming soon, and it is also arriving on Amazon Bedrock.
If you open ChatGPT today and do not see it, nothing is wrong with your account. That is just the schedule.
OpenAI is positioning Astra around computer use and software engineering rather than chat, which is a meaningful shift in framing. The company describes it as a new frontier in the speed, accuracy and safety of computer use, and as its best model for software engineering to date. On the benchmarks OpenAI published, Terminal-Bench 4.0 jumped from 37.3 percent to 57.9 percent, OSWorld 2.0 went from 65.7 to 72.6, ScreenSpot-Pro went from 76.9 to 92.7, and FrontierMath Tier 4 went from 83.0 to 97.6.
Standard caveat, and it is not a small one: every one of those numbers is OpenAI’s own, published at launch, with no independent replication yet. Treat them as a vendor’s claim about its own product, because that is exactly what they are.
The price is the story
API pricing is $10 per million input tokens and $50 per million output tokens. Cached input runs $1. Batch and flex processing are half price. A Fast mode costs double. Long context, meaning anything over 272,000 input tokens, jumps to $20 in and $75 out.
For comparison, GPT-5.6 Sol currently sits at $4 in and $20 out on a promotional rate that runs through at least November 21. So Astra is roughly 2.5 times the previous flagship on both sides of the meter.
Here is what that means in plain money. Say you run a customer support workflow that processes 2 million input tokens and generates 500,000 output tokens a month, which is a realistic load for a small team automating email triage and drafting replies. On Sol that is about $18 a month. On Astra it is $45. Not ruinous. Now scale that to a content operation running 20 million input and 5 million output tokens a month: $180 versus $450. The gap stops being academic somewhere around there.
And this is the part the launch coverage keeps skipping. Most small businesses do not touch the API at all. They use ChatGPT through a $20 or $25 per seat subscription, where Astra arrives at no extra cost as part of the plan. If that describes you, the pricing table above is trivia. You get the better model for the same money, and the only real question is whether it makes your work better.
The number that matters most to you
Buried under the benchmark table is the figure with the most direct bearing on how a small business actually uses this stuff. OpenAI reports a hallucination rate of 4.2 percent for Astra against 12.2 percent for Sol.
Think about what a 12 percent error rate means in practice. If you ask a model to pull ten facts out of a contract, more than one of them is wrong, and you have no idea which one. That is why so many owners tried AI for research or summarization, got burned once, and quietly went back to doing it themselves. The tax was verification, and verification ate the time savings.
Cutting that to roughly one in twenty-four does not eliminate the need to check. It does change the economics of checking. Spot-checking a mostly-right document is a different job from auditing a document you fundamentally do not trust.
If you abandoned AI-assisted research, summarization or data extraction because the output was too unreliable to be worth the review time, this is the release that justifies a second look. Not because the marketing says so, but because that specific number moved in the direction that was blocking you. Run your own test on your own documents before you believe it.
Computer use is the real upgrade
The benchmarks that moved most are the ones about operating a computer: clicking through interfaces, navigating web apps, completing multi-step tasks in a browser. OpenAI claims Astra completes those tasks 1.9 times faster than Sol on the Mind2Web benchmark.
For a small business, this is where the next two years of practical value sits. Not writing your blog posts. Logging into your supplier portal, pulling this month’s invoices, reconciling them against your accounting software and flagging the three that do not match. That kind of work is enormously tedious, extremely common, and has resisted automation because every small business’s stack is a slightly different mess of tools that were never designed to talk to each other.
A model that can reliably drive a browser does not need your tools to have an API. That is the unlock. It is also why we have been tracking the agent tooling space so closely, and why the gap between demo and dependable is the thing to watch.
Temper expectations accordingly. A 72.6 percent score on OSWorld means roughly one task in four still fails. You would not hand an employee with that record unsupervised access to your payment systems, and you should not hand it to a model either. Start with read-only work: research, data gathering, reporting. Save write access for when the numbers are much better.
The cyber gate, explained without the panic
Astra’s launch was delayed roughly four weeks because OpenAI’s internal evaluations put it past the company’s own Critical cybersecurity threshold. In practice that means the model can identify previously unknown vulnerabilities and build working exploits without step-by-step human guidance. During testing it reportedly found and chained two zero-days, which OpenAI disclosed to the affected maintainers.
The public version is trained to refuse advanced cyber tasks like writing proof-of-concept exploits. Looser safeguards are available to vetted organizations through a separate program that OpenAI says will expand.
For your business, the practical read is not that a model is going to attack you. It is that the capability now exists, is improving fast, and will not stay perfectly contained forever. The defensive posture that follows is boring and unchanged: patch things, use a password manager, turn on multi-factor authentication everywhere, and assume phishing is going to get better-written than it used to be. None of that is new advice. It is just more obviously worth doing.
What to actually do this week
If you use ChatGPT on a paid plan: do nothing except wait for it to show up, then re-run two or three tasks you gave up on because the output was not reliable enough. That is your test. If the answer is still wrong, nothing has changed for you and you can stop reading the launch coverage.
If you build on the API: do not migrate everything. The sensible pattern is routing. Keep cheap, high-volume, low-stakes work on a cheaper model and send only the hard tail to Astra, where the capability actually earns $50 per million output tokens. Anyone telling you to move your whole workload to the new flagship on launch week is describing a budget problem, not a strategy.
If you have not started with AI at all: this changes nothing about your first step. The bottleneck for a business that is not using AI yet has never been model quality. It is figuring out which of your repetitive tasks are worth automating. That work is the same today as it was last week, and it is where the actual leverage lives.
Everyone: resist the urge to rebuild anything. Frontier model launches arrive every few months now. The architecture decision that survives all of them is keeping your workflows loosely coupled to any one vendor, so switching is a config change rather than a project.
Frequently asked questions
Do I have to pay extra to get GPT-6 Astra?
Not on a ChatGPT subscription. It is rolling out to Plus, Pro, Business and Enterprise plans as part of what you already pay. The $10 and $50 per million token figures are API pricing, which applies to developers building applications, not to people using ChatGPT directly.
Is it actually better than what I am using now?
On OpenAI’s own benchmarks, yes, with the biggest gains on terminal tasks, computer use, math and cybersecurity. Those numbers have not been independently replicated. The honest answer is that you should test it on the work you actually do, because benchmark gains do not always translate into noticeable improvement on your specific tasks.
Should I cancel my other AI subscriptions?
Almost certainly not. Different models are better at different things, and a single frontier release does not collapse that. Audit your subscriptions because you should audit your subscriptions, not because of a launch announcement.
What does the 1 million token context window mean for me?
You can feed it much more material at once. Roughly speaking, a million tokens is somewhere around 750,000 words, so an entire contract library, a year of support tickets or a full codebase fits in one go. Note that going past 272,000 input tokens moves you to the higher long-context pricing tier if you are on the API.
Why does it cost 2.5 times more than the last model?
OpenAI has not published a cost breakdown. The plain reading is that a bigger, more capable model costs more to run, and that OpenAI is pricing to what the hardest agentic and engineering workloads will bear. The comparison is also slightly unflattering by design, since GPT-5.6 Sol’s $4 and $20 rates are promotional.
Related Coverage
- OpenAI’s Astra Solved 10 Open Math Problems for $2,000: our earlier look at the Astra line and what its research results implied for ordinary businesses.
- AI Agents Explained: the plain-language rundown of what agents can and cannot do, which is the capability Astra pushes hardest on.
- Best AI Writing Tools for Small Business: where the model underneath actually matters and where it does not.
Faceted Media Magazine covers business, AI, and entrepreneurship for the people building what’s next.