How to Get GPT-6 Astra: Access by Plan, the Free Question, and the Enterprise Toggle

The short answer, before the detail: GPT-6 Astra is usable today on ChatGPT Plus, Pro, Business and Enterprise, in ChatGPT Work and Codex, through the Open​AI API, and on Microsoft Azure and AWS Bedrock. It is not available on ChatGPT Free or Go. If you are on an enterprise workspace, it is off by default — an administrator has to enable it before anyone in the organisation can touch it. And the model itself is from September 3, 2026: this is a reference page for something that has finished rolling out, not launch coverage. Everything below was checked against Open​AI's own pages on September 16, 2026, and anything that comes only from rollout reporting is labelled as such.

The reason a model that is thirteen days old is worth a page this week is that the route in changed after the launch did. When Open​AI shipped GPT-6 Astra it described a rollout, not an availability: the model was going out "today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the Open​AI API, Microsoft Azure, and AWS Bedrock." Those coming days have now happened. Open​AI said on September 8 that Astra was "fully rolled out to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work"; as late as September 11 the company was still issuing paid subscribers one banked usage reset for each day they had gone without it; and by mid-September the model was being reported as broadly live across the ChatGPT plans, Azure and Bedrock. A reader asking how to get GPT-6 Astra in the first week of September was asking about a queue. Asking now, they are asking about a switch — and for enterprises, that is literally the answer.

Who can use GPT-6 Astra today, and what we verified

We read Open​AI's own availability line, model documentation and pricing page on September 16, 2026, and cross-checked the rollout state against reporting from the week before. Here is the split, stated as what it is.

• ChatGPT Plus, Pro, Business, Enterprise — named explicitly by Open​AI as receiving Astra. Vendor-stated, Open​AI's own launch page.

• ChatGPT Work and Codex — named as product surfaces carrying the model. This is the surface a Plus subscriber reaches Astra through. Vendor-stated.

• Open​AI API — live under the model id gpt-6-astra, priced and documented on Open​AI's own pages. Vendor-stated and live.

• Microsoft Azure and AWS Bedrock — named in the same availability line, with Microsoft Foundry reporting GA on Azure in the first week of September. Vendor-named; Azure GA date from third-party reporting.

• ChatGPT Free and Go — not named anywhere in Open​AI's availability line, and reported as excluded. Absence is the evidence: Open​AI has never listed a free tier for this model.

One wrinkle worth knowing about if you have been staring at a model picker. Rollout reporting through September says the Astra-powered option does not always present itself under that name: on Pro, Business and Enterprise accounts it has been reported as appearing in regular ChatGPT as "GPT-6 Pro", while Plus subscribers see it only in Work and Codex rather than ordinary chat, and Codex CLI needs version 0.153.0 or later. We could not confirm any of that on Open​AI's own documentation, which lists a single model id, gpt-6-astra, with no sibling tier. Treat the picker naming as reported, not vendor-documented — and note the practical consequence either way: if Astra is not in your picker, the model name is the wrong thing to be searching for.

Is GPT-6 Astra free?

No — and the honest version of that answer depends on which of three different questions you are actually asking, because "is GPT-6 Astra free" gets asked with at least three meanings behind it.

Free to use, if you already pay for an eligible plan? Yes, in the sense that matters day to day. Astra is included inside the usage allowances of the ChatGPT plans Open​AI names, rather than sold as an add-on you must buy before the model appears. What is not free is the plan. And the allowance is an allowance — reporting on the rollout describes Plus and Business Standard carrying consumption caps, with anything beyond purchased as credits, and describes rolling usage windows for the model. We could not verify specific message counts against Open​AI's own documentation and have not printed any.

Free at the API? No. Astra is a paid endpoint with no free tier attached. Open​AI's own documentation, read September 16, 2026, prices it at $10.00 per million input tokens, $1.00 cached input, $50.00 output at the standard tier for requests up to 272K input tokens; above that threshold the whole request reprices to $20.00 / $2.00 / $75.00. Batch and Flex run at half of standard and Fast mode at roughly double. There is no free trial credit for this model that we could find on Open​AI's pages, and we are not going to imply one exists.

Free as in open weights? No, and this is the one to say out loud because it quietly drives a lot of the searching. There is no downloadable checkpoint, no Hugging Face release, no self-host path. GPT-6 Astra is a closed hosted model on every surface it ships on, which means the only ways in are the ones listed above — a plan, or an API key, or a cloud marketplace.

So the truthful headline is: not free, and the cheapest way in is not the one most people assume. If you pay for an eligible ChatGPT plan you already have it, and the marginal cost of trying it is zero. If you do not, buying a subscription to run a handful of tasks is the expensive route — a per-token API call is the cheap one, because you pay for what you use and nothing on the days you do not.

The enterprise route: off by default, and an admin has to turn it on

This is the part of GPT-6 Astra access that catches people out, and it is the single most useful thing on this page for anyone in a company. Enterprise access is off by default at launch. An eligible workspace does not get Astra by virtue of paying for Enterprise; an administrator has to enable it under the organisation's applicable rate card and agreement. If a colleague tells you the company "has" Astra and you cannot see it, this is almost certainly why — check the workspace setting before you check the rollout.

What arrives with the switch is the genuinely enterprise-shaped half of the release. Open​AI's admin controls for Astra let an administrator restrict access to approved websites and desktop applications, manage uploads and downloads, and control browsing history — the guardrails you would want before pointing an agentic model at internal systems. ChatGPT Work and Codex add confirmation policies, which require approval before consequential actions, and automated review of unsafe or unauthorised tool calls. The design intent is visible in the shape of it: start narrow, widen as you learn.

Shipped alongside the model, in ChatGPT Desktop, were four enterprise plugins — Oracle Analytics, Power BI (a Microsoft Fabric service), Navan and Avalara. Their relevance to a page about access is that they are the practical reason an admin enables Astra at all: they are how the model reaches data and systems a team already uses rather than being a chatbot in a separate tab.

Two more things an enterprise buyer should know before signing anything. First, Zero Data Retention is available for eligible API customers on supported endpoints, subject to approval — eligible and subject to approval are doing real work in that sentence, so treat it as a process to start rather than a box that is ticked. Second, the enterprise gateway and the API gateway are separate: enabling Astra in a ChatGPT workspace does nothing for your API keys, and there is no admin toggle on the API side at all.

The API route, in one paragraph

Astra is live on the Open​AI API as gpt-6-astra, and the spec sheet is the reason people want it: a 1,050,000-token context window, a 128,000-token output ceiling, a knowledge cutoff of April 30, 2026, text and image input, reasoning-token support with effort levels from low to max, and the pricing above — all from Open​AI's own model documentation read September 16, 2026. One small discrepancy to flag rather than paper over: our own catalog listing for the model also shows file input alongside text and image, which is what a router records rather than what the vendor documents. If file input matters to your workload, check it against Open​AI's page before you build on it. There is no waitlist and no approval step for the standard endpoint; you need an account with billing, and the model is callable the same day. We have a separate walkthrough of the API surface, limits and request shapes if that is the route you are taking — this page is about getting in, not about what to send once you are.

What is not available

Stated plainly, because a page about access is only useful if it says where the doors are not.

• No waitlist, and no way to queue-jump. The initial limited group was selected by Open​AI rather than applied for. Nothing about the GPT-6 Astra rollout gave individuals a form to fill in or a position to hold, and on personal plans there is no mechanism to accelerate access.

• No free tier, and no free API credits. Covered above, repeated here because it is the breakout query and deserves to be findable.

• No open weights. No checkpoint, no local deployment, no third-party self-hosting.

• No country list, in either direction. Open​AI's availability line names plans and platforms, not geographies. We found no published regional restriction attached to this model — but equally, no promise that every region receives it on the same day. If Astra is missing from your picker and your plan, product surface and workspace settings all check out, availability by region is the remaining variable we cannot resolve for you.

• No offensive cyber capability, by design. This is the restriction most likely to surprise a technical reader. Astra is the first Open​AI model to reach the Critical cybersecurity threshold under the company's Preparedness Framework, and the shipped model refuses advanced cyber tasks such as writing proof-of-concept exploits. Open​AI says it plans to widen that with less restrictive safeguards for vetted defenders through its Daybreak programme — so the capability exists and is gated, rather than absent. For everyone outside Daybreak, the answer is that the model is available and that part of it is not.

• No verified free access via a third party. We saw aggregator and reseller listings for Astra during research, including endpoints quoted above and below Open​AI's own list price. Those are separate quotes on separate endpoints, not a discount on the vendor's rate, and we are not treating any of them as a route.

The cheapest way in, by who you are

Underneath all of this sits one question, and it has three different answers depending on which reader is asking.

If you are an individual. If you already pay for Plus or Pro, you have Astra already and the marginal cost of using it is nothing — look in ChatGPT Work and Codex rather than assuming the main chat picker is the whole product, and update your client if it is not there. If you are on Free or Go, the cheapest genuine route is the lowest paid tier Open​AI names, and the cheapest route if you only need Astra occasionally is not a subscription at all: a handful of tasks through the API costs a few dollars at $10/$50 per million tokens and costs nothing in the months you skip. Buying a monthly plan to run ten prompts is the most expensive way to do this.

If you are a team. Business is the plan that puts Astra inside a workspace you administer, which makes the admin controls and the confirmation policies the actual purchase rather than the model — you are buying the ability to let a team use an agentic model without giving it the run of your systems. Budget for credits above the included allowance rather than assuming the cap is the ceiling.

If you are an enterprise. The work starts before procurement: find out whether Astra is enabled in your workspace, because it is off by default and the answer is often that nobody has switched it on. Then treat it as two purchases — an enterprise agreement under your rate card with the admin controls, and separately, API access under a ZDR request if you need it. The first is an admin setting; the second is a process with an approval in it.

One cost note that matters whichever group you are in, because it changes the arithmetic of committing. Astra's list rate is $10.00 input and $50.00 output per million tokens, read from Open​AI's pricing page on September 16, 2026. GPT-5.6 Sol, the model it sits above, is billed on that same page on the same day at $4.00 / $20.00 — a promotional rate Open​AI says is available at least through November 21, 2026, and one that was cut from a pre-promotion $5.00 / $30.00 list. So Astra is 2.5x Sol at today's promotional rate on both axes, and 2x on input and about 1.67x on output against Sol's pre-promotion list. Older pages on this blog that print "2.5x" without naming which Sol rate they mean are quoting a number with an expiry date attached — the multiple belongs to Sol's discount, not to Astra's price. If you are deciding whether to commit a production path to a model at that premium, that is the number to budget against, and it is the kind of decision that is cheaper to keep reversible.

Which is the one place we earn a mention. GPT-6 Astra is in the OrcaRouter catalog as openai/gpt-6-astra, and OrcaRouter passes Open​AI's list rates through at 0% markup — so the vendor's price is our price, and a vendor move lands in your bill the same day rather than at renewal. That matters more than usual for this model, for a specific reason: Astra is expensive enough that most teams want to try it before they commit, and routing lets you put it behind the same key as the models you already run, send it a slice of real traffic, and fail over to something cheaper automatically when it is not earning the difference. The routing DSL composes several models into one call, one API covers 200-plus models, and switching traffic between them is a config change rather than a migration. Trying an unproven premium model is exactly the case that argument was built for.

Above is our own model page for the listing, which is where the routed pricing lives; below is Open​AI's documentation page for the model itself, which is the page to check when a spec question matters more than a cost question.

Where this leaves you

GPT-6 Astra access in one paragraph, as of September 16, 2026: paid ChatGPT tiers have it and Free and Go do not; enterprises have it and cannot use it until an administrator flips the switch under their rate card; the API has it at $10/$50 with no free tier and no approval step; Azure and Bedrock have it; there is no waitlist, no open weights and no region list, and the model's most dangerous capability is deliberately withheld behind Daybreak. The model is from September 3, 2026 — what is new this week is only that the "coming days" Open​AI wrote about on launch day are over, which turns a question about waiting into a question about choosing. For most readers that choice is small: check whether you already have it before you pay for it, and if you are going to pay for it, pay per token until you know it is worth the subscription.