Most model launches follow the same script. A company announces something, claims it is the best in the world, publishes a table where its bars are taller than everyone else's, and ships it the same day to anyone with a credit card.
GPT-6 Astra broke the last part of that script, and the break is more interesting than the model.
OpenAI announced it on 3 September 2026 as the most intelligent and aligned model in the world. It also delayed the launch by nearly a month, added safeguards, and shipped it with a portion of its own capability locked behind a vetting programme most businesses will never pass. A company that believed its product was simply safe would not have done that.
What actually shipped
The headline specification is a 1,050,000 token context window, which is roughly a million tokens and more than most businesses will ever deliberately use. Maximum output is 128,000 tokens, it accepts text and images, and the knowledge cutoff is 30 April 2026.
Pricing is $10 per million input tokens and $50 per million output. Cached input drops to $1, batch processing runs at half price, and there is a Fast mode at double the standard rate. That base rate is about 2.5 times GPT-5.6 Sol and twice Claude Opus 5, which puts Astra firmly at the expensive end of the market rather than in the steady deflation we have been tracking in what AI models actually cost now.
The rollout was staged rather than simultaneous. A limited set of organisations got access on day one, followed by ChatGPT Plus, Pro, Business, and Enterprise over subsequent days, plus the OpenAI API and AWS. General availability landed on 4 September. If you are on a paid ChatGPT plan, you have it.
The part they locked away
This is the genuinely unusual part of the launch and it deserves more attention than the benchmark table.
OpenAI's internal evaluations put GPT-6 Astra past its own Critical cybersecurity threshold, meaning the model can find and exploit software vulnerabilities autonomously. Not assist a security researcher. Find and exploit, without a human driving. That finding caused OpenAI to slow the release on 7 August, add safeguards, and ship the advanced cyber capabilities restricted to vetted organisations through a trusted-access programme.
Read alongside the Anthropic incidents, a pattern emerges that is worth naming plainly. The frontier labs are now building models whose full capability they are not comfortable releasing openly, and they are saying so in public documentation rather than being caught. That is a meaningful improvement in candour and a fairly stark statement about where capability has got to.
For a small business the direct impact is close to zero, because you will never be asked whether you want the gated tier and you would have no use for it. The indirect impact is worth understanding: the same capability that is being withheld from you is the capability that exists in the world, and the people most interested in autonomous vulnerability discovery are not applying through a vetting programme. The defensive implication is the one that reaches your business, and it looks like the ordinary security hygiene we went through in the AI agent security guidance.
What the benchmarks actually show
The numbers are strong and worth reading carefully rather than as a scoreboard, because one of them is far more relevant to you than the rest.
On computer use, meaning the model operating software the way a person would, Astra scores 72.6% at roughly 40 minutes per task, which is about 47% faster than GPT-5.6 Sol. On Agent's Last Exam it reaches 59.3%, against 48.7% for Claude Fable 5 and 55.5% for Claude Opus 5. Those are real margins rather than rounding, though the gap over Opus 5 is narrower than the headline framing suggests.
The number that actually matters arrived quietly. On an evaluation built specifically after the Hugging Face sandbox incident, measuring how often a model takes actions outside its authorised scope, Astra scored 0%. GPT-5.6 Sol, running without production safeguards, scored 48.2%. That is not a capability measurement, it is a containment measurement, and given that the last two months of AI news have been dominated by models doing things outside their intended scope, it is arguably the most important figure OpenAI published.
The caveat applies as always: this is a vendor evaluating its own model on a benchmark it built, and 0% on a constructed evaluation is not the same as 0% in your business. But the fact that out-of-scope action is now something labs benchmark and publish at all is a change from a year ago, and it is the metric to watch as these evaluations mature.
The price is the real decision
For a small business, everything above is context and this section is the decision.
At $10 and $50 per million tokens, Astra costs roughly two to five times what the models you are probably already using cost. Claude Opus 5 sits at $5 and $25. Claude Fable 5.1, released two days before Astra, is cheaper again for typical work. Grok 4.5 runs at $2 and $6. The question is not whether Astra is better, because on most measures it is. The question is whether it is better at the specific thing you do, by enough to justify the multiple.
For the overwhelming majority of small business AI work, the honest answer is no. Classifying support tickets, drafting emails, extracting data from invoices, and answering customer questions are tasks where the difference between a frontier model and a mid-tier one is invisible in the output and very visible on the invoice. Paying frontier prices for classification is the most common way businesses overspend on AI, and it has been true through every model generation.
Where the multiple might be justified is long-running autonomous work, where the computer-use and agent benchmarks translate into fewer failed runs. A task that completes successfully at Astra prices beats a task that fails twice at cheaper ones, and the 47% speed improvement on computer-use tasks is a real operational difference if that is what you do all day. The cached input rate of $1 also matters more than it looks for anything that repeatedly sends the same large context, which is a common shape for document-heavy workflows.
Who actually needs a million tokens
A million tokens is roughly 750,000 words, or about nine copies of a long novel. It is worth being concrete about that because context window numbers have become a marketing arms race disconnected from use.
Almost no small business task needs it. A long customer email thread is a few thousand tokens. An entire year of your support tickets might be a few hundred thousand. The genuine use cases are narrow: analysing a large contract set in one pass, working across an entire codebase, or processing a lengthy document collection where splitting it would lose cross-references that matter.
The trap is that large context windows invite lazy prompting, and lazy prompting is expensive. If a million tokens are available, the temptation is to send everything rather than the relevant part, and you are billed for everything you send. A workflow that sends 200,000 tokens of context to answer a question that needed 2,000 is paying a hundred times too much, and it will produce a worse answer too, because relevant information gets diluted.
The discipline that saves money here is the same one that improves quality: send what matters. That is unglamorous advice and it outperforms model selection in most small business workflows, which is why we keep returning to it in pieces like routing between models to cut AI costs.
The honest verdict
GPT-6 Astra is a genuinely strong model and it is not a reason for most small businesses to change anything this month.
If you use ChatGPT on a paid plan, you already have it and you should simply use it, because on the chat interface the model upgrade costs you nothing extra and the capability is real. If you call models through an API for production work, evaluate it only against the specific workflows where you are currently seeing failures, and expect most of your volume to remain on cheaper models where the difference does not show.
The part of this launch worth remembering in six months is not the benchmark table. It is that OpenAI tested its own model, concluded part of it was too capable to release openly, delayed the launch, and said so publicly. Whether that becomes the norm or a one-off is the more consequential question, and it will matter to your business long after the price of a million tokens has fallen again.
Sources
- MarkTechPost: OpenAI releases GPT-6 Astra, a 1.05M-context computer-use model gated behind a Critical cyber threshold
- 9to5Mac: OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra
- LLM Stats: GPT-6 Astra benchmarks, pricing and context window
- Digital Applied: GPT-6 Astra price, access and what the benchmarks show