GPT-6 Astra Explained: Which ChatGPT Plan Actually Gets It, and Whether You Should Upgrade
What's in this guide
OpenAI released GPT-6 Astra on September 3, 2026, two days after Anthropic shipped Claude Fable 5.1. If you pay $20 a month for ChatGPT Plus and opened the app looking for it, you already know the frustrating part: it is not there, and the launch post does not really explain why.
So here is the plain English version, checked against OpenAI's pricing and help center pages on September 4: what Astra is, why ChatGPT is now really three products, which plan gets the new model and where, what it costs, and a two minute decision on whether any of it is worth your money.
The short answer
GPT-6 Astra is OpenAI's new top model, built for long jobs that involve using a computer: filling in forms, building spreadsheets and decks, writing and testing code. In the normal chat window it appears as GPT-6 Pro, only on the Pro plans ($100 or $200 a month), with weekly caps. Plus subscribers do not get it in Chat at all, only a limited amount inside ChatGPT Work and Codex, the agent modes, with credits for more. Free and Go get nothing. If you mostly ask ChatGPT questions, this launch should not change what you pay.
What GPT-6 Astra actually is
Astra is the successor to GPT-5.6 Sol, the model paying ChatGPT users have been talking to since July. OpenAI's pitch is not "it chats better." The pitch is that it can do work on a computer the way a person does: open apps, click through websites, build a deck from your template and check the result before handing it back.
OpenAI's own numbers back that up in a narrow way. On OSWorld 2.0, a test of real desktop tasks, Astra completed 72.6 percent in about 40 minutes each, against 65.7 percent in about 75 minutes for GPT-5.6 Sol. The launch demos say it all: booking a DMV appointment, apartment hunting, filling in a Form 1040.
Two independent checks matter more. Artificial Analysis scored Astra at 61 on its general Intelligence Index, level with GPT-5.6 Sol and five points behind Claude Fable 5.1, though it hallucinated on roughly half as many questions as Sol. And the ARC Prize team, whose ARC-AGI-3 test Astra "saturated" at 99.9 percent, noted that the score used OpenAI's own harness. With the standard one it scored 62.7 percent. A record, not the AGI moment the headline implies.
One more thing: Astra is the first OpenAI model to hit the "Critical" cybersecurity level in the company's risk framework. That is why security customers got it first, why it refuses to write hacking tools even for defenders, and why OpenAI's monitoring can pause a task in ChatGPT or Codex and ask you to review it. Expect the occasional interruption.
Chat, Work and Codex: the three ChatGPTs you need to know first
Here is what makes every "do I get Astra" question confusing. Since July, ChatGPT has been three products wearing one subscription, and OpenAI's help pages assume you know which is which.
Chat is the familiar window. On paid plans it runs GPT-5.6 Sol with a slider for how hard it thinks (Instant, Medium, High, Extra High, Pro). Free and Go get the cheaper GPT-5.6 Luna.
Work is the agent. You describe an outcome (a report, a spreadsheet, a deck, a website) and it goes off and builds it, running code, browsing and asking questions along the way. It runs in the cloud on web and mobile, and on your own computer through the desktop app. Work is for paying subscribers. Free and Go get a limited taster on desktop only.
Codex is the coding agent, in the desktop app, the command line and code editors. The important fact for this post is that Work and Codex share one usage allowance, separate from Chat's. Once you have that picture, the plan rules make sense.
Which ChatGPT plan gets GPT-6 Astra
This is straight from OpenAI's help center articles on GPT-6 Pro and on ChatGPT Work and Codex, both updated this week.
Free and Go (free, or $8 a month)
No Astra anywhere. You keep GPT-5.6 Luna in Chat and the limited desktop taste of Work.
Plus ($20 a month)
Not in Chat. OpenAI's wording is blunt: GPT-6 Pro "is not included with ChatGPT Plus in Chat." Your model picker stays on the GPT-5.6 family, with Medium and High reasoning but not Extra High or Pro.
Where Plus does get Astra is inside Work and Codex. OpenAI calls it "limited Astra usage," drawn from your normal allowance, with credits available once that runs out. The Codex pricing page puts a rough number on it: 5 to 45 Astra messages in a five hour window, against 10 to 100 for GPT-5.6 Sol. Treat it as something you switch to for one hard job, not the default.
Pro ($100 or $200 a month)
Both Pro tiers get Astra in Chat, listed as GPT-6 Pro in the model picker, and in Work and Codex on the plan's full allowance.
The Chat caps are the part to read twice. On the $200 plan you get 200 GPT-6 Pro messages a week, then ChatGPT drops you back to GPT-5.6 at Medium. On the $100 plan you get 50 a week, shared with the older GPT-5.6 Sol Pro, so switching between them buys nothing. A Pro subscription does not mean unlimited Astra.
In Work and Codex, Pro $100 is about 5x the Plus allowance and Pro $200 about 20x. OpenAI's Astra estimates are 25 to 225 messages per five hours on the $100 tier and 100 to 900 on the $200.
One rollout warning
At the time of writing Astra was live only for a small set of enterprise security customers, with Plus and Pro access arriving "in the coming days." If you do not see it yet, that is why. Buying credits does not get you in early. OpenAI says so explicitly.
What it costs once the included usage runs out
Plan prices. Free is $0. Go is $8 a month and may show ads. Plus is $20. Pro is $100 for 5x the Plus usage or $200 for 20x. No personal plan offers annual billing. All checked on September 4.
Credits in Work and Codex. When a Plus or Pro user runs out of included allowance, extra usage is billed in credits. Astra costs 250 credits per million input tokens and 1,250 per million output. GPT-5.6 Sol costs 100 and 500. That is 2.5x, and OpenAI warns separately that Astra "can use your allowance faster" before you touch credits at all. Credits work only in Work, Codex and the Excel add-in, never in Chat, and expire after 12 months.
The raw API. For developers, Astra is $10 per million input tokens and $50 per million output, exactly what Anthropic charges for Fable 5.1 and 2.5 times what GPT-5.6 Sol cost. Fast mode doubles the speed and the price.
If those numbers make you wince, a capable model running on your own computer costs nothing per message. Different tool, different job, but it is the honest comparison for anyone whose ChatGPT use is mostly questions and drafts.
What I found on day one
I am a Plus subscriber, so the first thing I did on September 4 was look for Astra in my own account. Chat: nothing. The model picker still reads GPT-5.6 with Instant, Medium and High, and no GPT-6 entry, greyed out or otherwise. Work: Sol, Terra and Luna, no Astra. That matches the "coming days" line, and it tells you how much of the launch day excitement applies to a normal subscriber. None of it, yet.
I did get about an hour with Astra through the API, on the kind of task OpenAI is selling it for: a folder of a dozen PDFs, turned into a one page comparison table and a short recommendation. Two things stood out. It asked one clarifying question before starting (which of two overlapping columns I wanted), then got on with it, a change from Sol's habit of guessing. And the output was about a third shorter for the same content, which fits OpenAI's claim that Astra pulls in only the context that matters.
The flip side is that at $10 and $50 per million tokens, the hour cost more than a week of Sol would have. For a Plus user that arithmetic is the whole story.
Astra vs Claude Fable 5.1: the $20 question
Both launched in the same week, so the natural question is which $20 plan gets you the better top model. Neither company gives its best model to $20 subscribers in the normal chat, but they take different roads.
Anthropic's approach with Fable 5.1 is pay as you go: on the $20 Claude Pro plan you turn on usage credits and pay API rates per message, with a cap you set. On the $100 Max plan it is included up to half your weekly limit. OpenAI gates by product instead: nothing in Chat on Plus, some Astra inside Work and Codex, and on Pro a capped amount in Chat plus your full allowance in Work and Codex.
If you want the top model for the occasional hard question in a chat window, Claude Pro plus a small credit balance is the cheaper route. If you want an agent that goes away and builds things, ChatGPT Plus already includes a taste of Astra in Work for no extra money, which is more than Anthropic gives away at that price. On raw ability the independent scores put Fable 5.1 slightly ahead on reasoning and Astra ahead on computer use, a gap most people will never notice next to the billing differences. If you are still deciding which assistant to use at all, my Claude vs ChatGPT vs Gemini comparison is the wider picture.
Who should upgrade
Here is the decision as a chart, and then in words.
Stay where you are if ChatGPT is where you ask questions, draft emails and get explanations. Astra never appears in Chat on Plus, so upgrading for it makes no sense, and GPT-5.6 Sol is more than good enough for that work.
Stay on Plus and try Astra in Work if you occasionally hand ChatGPT a real job: a report from a pile of files, a spreadsheet, a deck, a script. Your included allowance covers a handful of Astra runs per session. Use Sol or Terra for routine tasks and save Astra for the one that matters. Buy credits only if you keep wanting more, and set a spend limit.
Go to Pro $100 if you already hit the Work and Codex limit on Plus most weeks. The 5x allowance is the real reason to upgrade. The 50 GPT-6 Pro messages a week in Chat are a bonus, not the point.
Go to Pro $200 only if you run Work or Codex for hours a day. Astra burns allowance about twice as fast as Sol, so "20x" is closer to 10x once you switch models.
Do not upgrade because of benchmarks. Independent scores put Astra roughly level with the model you already have on general questions. Its gains are in long, hands-on computer tasks, and if you do not give it those, you will not see them.
The verdict
GPT-6 Astra is a real step forward for one specific thing: letting an AI take over a long, fiddly computer job and finish it. OpenAI has priced it accordingly and drawn a hard line between Chat, where Plus users will not see it, and Work and Codex, where they get a taste. The Pro plans get it properly, with weekly caps you should read before you pay.
My recommendation: if you are on Plus, wait for the rollout to reach you, then try Astra inside Work on your next hard task and see whether the result is worth the allowance it burns. If you keep reaching for it and keep running out, Pro $100 is the sensible next step. If you never reach for it, you have lost nothing, and the ChatGPT you already have got quietly faster this week anyway.
Comments
Post a Comment