How to Access and Use GPT-6 Astra

If you pay for ChatGPT Plus, opened the model picker expecting to find GPT-6 Astra and it was not there, nothing is wrong with your account. On Plus, Astra lives in ChatGPT Work and in Codex, not in the normal chat window. The model you see offered inside regular chat is called GPT-6 Pro, and that one needs a Pro, Business or Enterprise plan.

That single distinction explains most of the confusion around this launch, and it is why several of the pages ranking for this question right now are forum threads rather than guides. This article covers where GPT-6 Astra actually appears on each plan, what you are allowed to use, how to reach it through the API, and what it is genuinely good at.

Everything here reflects the state of the rollout as of 24 September 2026. This is a three week old product and OpenAI has been changing the arrangement week to week, so check your own model picker against the table below rather than assuming it is still current.

What GPT-6 Astra is

GPT-6 Astra is OpenAI’s frontier model, released as a limited preview on 3 September 2026 and to paying users generally the following day. It replaces GPT-5.6 at the top of the range. OpenAI describes it as state of the art in coding, mathematics, and operating computers and web browsers, and pitches it at long running agentic work rather than at conversation.

The company has been unusually loud about it. President Greg Brockman suggested it may eventually be seen as the arrival of artificial general intelligence, and OpenAI says the training run was its largest ever, the first pretrained on more than 100,000 GPUs at its Stargate site in Texas. Treat the AGI framing as marketing and the engineering claims as checkable.

Which plans get GPT-6 Astra, and where it appears

This is the table to read before anything else, because the model name changes depending on where you meet it.

Plan	Do you get Astra?	Where it appears
Free	No	Not available
Go	No	Not available
Plus	Yes	ChatGPT Work and Codex, not normal chat
Pro $100	Yes	GPT-6 Pro in chat, plus Work and Codex
Pro $200	Yes	GPT-6 Pro in chat, plus Work and Codex
Business	Yes	GPT-6 Pro in chat, plus Work and Codex
Enterprise	Yes	GPT-6 Pro in chat, plus Work and Codex
API	Yes	Model id gpt-6-astra, no plan tier

So there are two different things wearing similar names. GPT-6 Astra is the model. GPT-6 Pro is how the higher reasoning version of it is presented inside the ChatGPT chat interface, and that surface is gated to Pro and above. Plus subscribers have the model, they just reach it through a different door.

OpenAI’s launch post said Astra would reach all Plus, Pro, Business and Enterprise users, and the plan documentation then described Plus access as being in Work and Codex. Sam Altman called the rollout messy on 4 September, and paying users were given one banked usage reset for each day they went without the model, counted from 3 September. As of today, no OpenAI staff member has publicly settled whether normal chat access for Plus is coming or permanent.

How to access GPT-6 Astra on ChatGPT Plus

1. Open ChatGPT and look for the Work surface rather than your usual chat. This is a separate space with its own conversation history.
2. Open the model picker inside it. Astra appears there if the rollout has reached your account.
3. If you would rather work in the terminal or an editor, use Codex, where Astra is also available on Plus.
4. If it is in neither, you are waiting on the rollout rather than doing something wrong. There is no setting to switch it on.
5. Do not go looking for it in the normal chat model picker. On Plus it is not there, and the highest model you will see in chat is GPT-5.6 Sol.

Worth saying plainly, because it costs people money: upgrading from Plus to Pro to get Astra in normal chat buys you the chat surface and a larger allowance, not access to a model you did not already have.

How much you can actually use

Astra is expensive to run and the allowances reflect that. These are the published figures, and OpenAI notes that Astra burns an allowance faster than GPT-5.6 Sol depending on task complexity, context length and reasoning settings.

Plan	Allowance	Period
Plus	Roughly 5 to 45 Astra messages	Per 5 hours
Pro $100	50 GPT-6 Pro messages, shared with GPT-5.6 Sol Pro	Per week
Pro $200	200 GPT-6 Pro messages	Per week
Business Standard	15 Pro messages	Per month

The Plus range is wide for a reason. Five to forty five is not vagueness, it is the model metering by work done rather than by message sent. A one line question and a forty step browser task are not the same unit, which is worth remembering before you spend your window on something trivial.

How to use GPT-6 Astra through the API

The API has no tier gate. If you have an OpenAI API account you can call it today, and this is the route worth taking if the ChatGPT allowances are the thing getting in your way.

Specification	Value
Model id	gpt-6-astra
Context window	1,050,000 tokens
Max output	128,000 tokens
Input price	$10 per million tokens
Output price	$50 per million tokens
Cached input	$1 per million tokens
Cache writes	$12.50 per million tokens
Knowledge cutoff	30 April 2026
Input types	Text and image
Output types	Text

Two numbers in that table deserve attention. The context window is just over a million tokens, which changes what is worth attempting: whole repositories and long document sets fit in one pass. And cached input at $1 against $10 for fresh input is a tenfold difference, so if you are sending the same long prefix repeatedly, structuring for prompt caching is the single largest saving available to you.

Full parameter details are in the OpenAI model documentation for gpt-6-astra, which is the one source worth checking directly, because pricing and limits on a new model move.

The reasoning effort setting, and why it matters to your bill

Astra exposes a reasoning.effort parameter with five settings: low, medium, high, xhigh and max. It controls how much thinking the model does before answering.

This is the main cost and latency dial on the model. Reasoning tokens are billed at the output rate, so raising effort raises the bill on the same prompt, sometimes considerably. The useful habit is to start low, and only raise it when you can point at a specific failure that more deliberation would fix. Most straightforward extraction, classification and drafting work does not improve at high effort, it just costs more.

What it is actually good at

OpenAI positions Astra around long horizon agentic work, meaning tasks with many steps that run over time rather than single questions. The tool list bears that out. Alongside the usual web search and file search it supports a code interpreter, a hosted shell, patch application, computer use, MCP connections and tool search.

• Multi step computer and browser work, which is the flagship claim. OpenAI’s own examples include filling in tax returns, ordering food and running job searches.
• Software engineering across a whole codebase, helped considerably by the million token context window.
• Deep research and scientific work, where the reasoning effort dial earns its keep.
• Document creation and long analysis, the routine professional workload.

On benchmarks, the number OpenAI leads with is ARC-AGI-3, where it says Astra beat the human action efficiency baseline on 96% of levels. That is a real result on a hard benchmark. It is also a benchmark about efficient action in novel environments, not a general claim about the model being right, so read it as evidence about agentic competence specifically.

What it will refuse, and the safety context

Astra is the first OpenAI model to reach the Critical level for cybersecurity capability under the company’s Preparedness Framework. The practical consequence is that the publicly released version is deliberately restricted: it declines advanced cybersecurity work, including writing proof of concept exploits for vulnerabilities. If you do legitimate security work, expect refusals on tasks earlier models would attempt.

There is a second thing worth knowing, because it is unusual. Astra uses a reasoning technique variously called recurrent depth or looped transformers, which is more efficient but obscures some or all of the model’s chain of thought. Outside researchers have raised concerns about monitorability as a result, and OpenAI’s chief scientist Jakub Pachocki has said preventing unintended harm is getting harder and may bottleneck further progress.

None of that stops you using it. It does mean that for consequential work you should be checking the output rather than the reasoning, because on this model the reasoning is less available to check than it used to be.

Should you upgrade for it?

If you	Then
Are on Free or Go	No access at all, upgrading to Plus is the entry point
Are on Plus and want agentic work	You already have it, in Work and Codex
Are on Plus and want it in normal chat	That needs Pro, and buys a surface, not a model
Hit the Plus 5 hour window constantly	The API is usually cheaper than a Pro upgrade
Are building a product on it	API only, there is no tier gate there
Mostly write and summarise	GPT-5.6 Sol is fine, Astra is aimed elsewhere

That last row matters more than the marketing suggests. Astra is built for multi step work with tools. If your use is drafting, summarising and answering questions, the older model does that without consuming a scarce allowance.

Quick answers

For more explainers on how these tools actually work, see our Tech section, or our guide to what agentic AI can and cannot do, which covers the loop Astra is built to run.