Quick answer: Grok 4.7 came out on September 21, 2026, at exactly Grok 4.6's price: $2 per million input tokens, $6 per million output. It is a real but small step up. xAI's own coding scores rise 6 to 17 points, and Artificial Analysis moves its overall index just 2.
xAI launched it in Cursor, Grok Build and the API. The launch page never mentions the Grok app or SuperGrok.
My verdict after running it in Grok Build: same price, slightly better. I'm like, cool. But these are minor updates. And the first job I gave it failed.
My read on Grok 4.7: cheap per token, hungry per job, third in agentic coding by its own founder's count.
Published September 22, 2026, the day after launch. Prices, benchmarks and availability were checked on xAI's, GitHub's and Artificial Analysis's own pages that day.
Grok 4.7 at a glance: same price, small step upQuestion
AnswerRelease date
September 21, 2026 (xAI's launch post)Price
$2 in / $6 out per 1M tokens, same as 4.6Context window
500,000 tokens (xAI's docs)Real step up from 4.6?
Small. Coding scores up 6 to 17 points; overall index up 2Where xAI launched it
Cursor, Grok Build, the Grok API, coding tools, model routers, clouds; GitHub Copilot rolling outIn the Grok app / SuperGrok?
Not on xAI's launch list. Signed out, grok.com's picker shows modes, not model numbersGrok 4.7 Fast
Same model, twice the speed, twice the price; Cursor and Grok Build onlyMy verdict
"Same price, slightly better." A minor updatexAI's launch page, September 22, 2026. Last line: "Served at the same price and speed as Grok 4.6."
Grok 4.7 shipped to Build, Cursor and the API first
Here is where xAI says Grok 4.7 runs, word for word from the launch post: "Grok 4.7 is available today in Cursor and Grok Build. It is also available through the Grok API, third-party coding harnesses, and model routers and cloud platforms."
Read that list again. There is no Grok app in it, no grok.com, no SuperGrok, no Heavy.Surface
Has Grok 4.7?
SourceGrok Build (coding agent)
Yes, the default model per xAI's docs; chat-window Build mode showed 4.7 on my account
xAI's launch post and docsCursor
Yes, on all Cursor plans
xAI's docsGrok API
Yes, model id grok-4.7
xAI's docsOpenRouter, Vercel, Cloudflare
Yes
xAI's docsGitHub Copilot
Rolling out gradually, paid plans only
GitHub's changelogGrok app chat (grok.com, SuperGrok, Heavy)
Not listed. The signed-out picker shows no model numbers
xAI's launch post; screenshot belowGrok Bot
Unknown. It will not say which model it runs
Grok Bot itself, see belowgrok.com signed out, September 22, 2026. The picker shows modes, not models. Nothing says 4.6 or 4.7.
GitHub's wording is "Rollout will be gradual," for Copilot Pro, Pro+, Max, Business and Enterprise. So if your Copilot picker does not show it yet, that is expected.
Grok 4.7 Fast is the one to watch on your plan. Per xAI's docs it is "the same model served on faster infrastructure, billed at twice the standard token rates," it lives only in Cursor and Grok Build, and "it is not included in Grok Build's free tier."
My first job for Grok 4.7 in Build failed
I have been running Grok 4.7 in Grok Build. We confirmed the 4.7 label in Build, so that is the one surface where I know which model I am on.
The first thing I tried was having Grok Build make an app, or a script, to help me find good business ideas. Just to test it. It failed.
For what it is worth, this is something Codex flunked too, earlier. So 4.7 did not clear a bar that Codex had already missed.
That is honestly all I have tested it on so far. We have been building, or at least failed to build, and that is the whole field test. There is no log of that run on my machines, so I am not going to describe the error or pretend I timed it.
If Build mode itself is new to you, start with what Grok Build actually is. If you are choosing between it and Claude's version, that is Grok Build vs Claude Code.
Grok 4.7 vs 4.6: coding up 6 to 17 points, overall up 2
My honest reaction to the upgrade: "Same price, slightly better. I'm like cool. But it's just like these are minor updates, you know what I mean?"
The numbers say the same thing, with one surprise in the coding rows.Measure
Grok 4.6
Grok 4.7
ChangePrice per 1M tokens (in / out)
$2 / $6
$2 / $6
NoneCursorBench 4.0 (xAI)
40.4%
46.3%
+5.9 pointsTerminal-Bench 4.0 (xAI)
20.3%
37.6%
+17.3 pointsDeepSWE v1.1 (xAI)
65.2%
71.0%
+5.8 pointsGDPval, office-work Elo (xAI)
1605
1695
+90Artificial Analysis Intelligence Index
44
46
+2Artificial Analysis Coding Agent Index (with Grok Build)
47
56
+9Output tokens per index task (Artificial Analysis)
~36k (high)
~81k (xhigh)
More than doublexAI's numbers are xAI's own runs. The Artificial Analysis write-up is the independent check, and it agrees on the shape: a big jump on coding agents, a nudge on everything else.
Compare that to the same week at Anthropic. Claude Opus 5.5 landed one day later, and Anthropic says it "performs at the level of Claude Fable 5.1 on most work" at $4 / $20 per million tokens. That is the bigger update. My full take is in Opus 5.5 vs Fable.
Grok 4.7 is cheap per token, not cheap per job
When I first talked this through, I called 4.7 "way more token efficient." That is half right, and the half that is wrong matters for your bill.
Per token, it is cheap: $6 per million output tokens, against $20 for Claude Opus 5.5 and $50 for Claude Fable 5.1. Per job, it is hungry. Artificial Analysis measured about 81,000 output tokens per task at xhigh effort, against 36,000 for Grok 4.6 and 27,000 for GPT-6 Astra at max.
So the efficiency play is real. It is just the price tag doing the work, not a model that says less.
A Hacker News user, mchusma, noticed the same thing on day one: "It feels kind of like they really had it burn tokens to claw up the benchmarks."
To be fair to xAI, its own CursorBench chart plots Grok 4.7's average cost per task at roughly half of Fable 5.1's or Opus 5's, which is where the launch tagline "half the price of comparable models" comes from. Cheap tokens can still beat expensive ones even when you burn more of them.
Third place, by Musk's own count
What I heard going in was that 4.7 would be at Opus 5 level, and it turned out worse on some things. That is close to what Elon Musk actually posted a week before launch.Musk's pre-launch post, September 14, 2026: 4.7 "roughly on par with Opus 5.0, not 5.1."
After launch he was blunter. His September 21 post, via the Musk post archive: "Grok 4.7 places @SpaceXAI as third, after Anthropic & OpenAI, for agentic coding. When factoring in that Grok is significantly faster & lower cost, it's a great choice for your everyday workhorse."
Artificial Analysis puts it one spot lower. Grok 4.7 with Grok Build ranks fourth on its Coding Agent Index, behind Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.
One correction to my own memory. I thought Elon said Grok 5 would be at Fable level. His September 14 post actually says "Grok 4.9 is probably Astra/Fable class" and "Grok 5 maybe better than anything." Either way, I am waiting on Grok 5.
Because here is the part I actually care about: if you can get Fable-like intelligence at Grok prices, that becomes very intriguing.
Where Grok genuinely wins: price, speed, and a free Build door
I do not think xAI's models are impressive right now. I also think Grok is not to be messed with at this point. They are playing the efficiency game well. 4.6 was pretty good, and 4.7 is solid, for what it is worth.
Here is what Grok 4.7 has going for it:Price per token. $2 / $6 per million, against $4 / $20 for Claude Opus 5.5 and $10 / $50 for Claude Fable 5.1. The output rate is under a third of Opus 5.5's.
More model, same bill. moojacob on Hacker News: "Apparently Grok 4.7 has 40% more weights than Grok 4.6, but the price ($6 output token, $2 input) is the same." xAI has not published a parameter count, so treat the 40% as reporting, not spec.
Speed. xAI serves it at 4.6's speed, and Grok 4.7 Fast gives you 2x the output speed at 2x the price. Musk's own pitch is "significantly faster & lower cost."
A free door. Grok Build has been on every plan, including Free, since August 19, 2026.Grok Bot will not tell you if it runs 4.7
Grok Bot ironically can't confirm if it's using 4.7, so I can't even tell you.
My guess is that a day after launch, it is. But nobody knows what the Bot uses under the hood; it just does whatever it does. So I am not going to vouch for 4.7 there. Build is the only place I have confirmed it.
If you are weighing the Bot against ChatGPT's agent, that comparison lives in GPT-6 Astra vs Grok Bot.
What to do about Grok 4.7 this week
Nobody should change plans over this release. Here is the move by who you are:If you use free Grok, open Build and try it. Grok Build is on the free plan, and xAI's docs say the Fast variant is held back from Build's free tier.
If you pay for SuperGrok, you already have 4.7 where xAI put it, in Grok Build. In regular chat, xAI has not said, and I have not captured my signed-in picker yet. Build and chat draw from one weekly pool, so read how Grok's usage limits work before you point long Build jobs at it.
If you pay for Claude or ChatGPT, this is not the week to switch. The same week brought Claude Opus 5.5, the bigger update in my book, and GPT-6 Sol. For everyday work, my setup is on Claude guides.
If you code in Cursor, this is where 4.7 makes the most sense: it is on every Cursor plan, cheap per token, and fourth on Artificial Analysis's coding-agent ranking. Try it on a job where speed matters more than the last few points of quality.FAQ
Is Grok 4.7 free?
Partly. Grok Build has been on every Grok plan including Free since August 19, 2026, and xAI's docs say the faster Grok 4.7 Fast variant is not included in Grok Build's free tier. xAI's docs describe the Grok Build coding agent; they do not say which model the chat window's Build mode (the hammer icon) runs on a free account.
Through the API you pay per token: $2 per million input tokens and $6 per million output tokens.
Is Grok 4.7 in SuperGrok?
In Grok Build, yes: xAI launched it there, Build is on every plan, and Build showed 4.7 on my SuperGrok Plus account. For regular Grok chat, xAI has not said. Its launch post never mentions the Grok app, SuperGrok or Heavy, and the signed-out grok.com picker shows modes, not model numbers.
What is the difference between Grok 4.7 and Grok 4.6?
Same price ($2 in, $6 out per million tokens), better coding scores. On xAI's own table CursorBench 4.0 goes from 40.4% to 46.3% and Terminal-Bench 4.0 from 20.3% to 37.6%.
Artificial Analysis moves it only 2 points on its overall Intelligence Index (44 to 46) but 9 points on its Coding Agent Index (47 to 56).
What is Grok 4.7 Fast?
The same Grok 4.7 model served on faster infrastructure, at twice the token price. xAI's docs say it is available only in Cursor and Grok Build, is not included in Grok Build's free tier, and is not on the public xAI API.
Is Grok 4.8 next?
Probably, but nothing is published. On September 14, 2026, Elon Musk posted that Grok 4.8 will be a noticeable improvement and that Grok 4.9 is probably Astra/Fable class. There is no release date, model card, or price for either.
Where this comes from
The facts come from xAI, GitHub, Artificial Analysis and Musk's own posts, all read September 22, 2026. What's mine is my SuperGrok Plus account and one failed Grok Build run on 4.7.
I haven't run my own benchmark, and I don't have a screenshot of my signed-in Grok app yet. Last reviewed September 2026.
Quick answer: GPT-6 Sol and Luna are OpenAI's two cheaper GPT-6 models, released September 22, 2026, sitting under GPT-6 Astra.
Three names, one ladder, one rule: Astra for the hard jobs, Sol for everyday work, Luna for high-volume grunt work.
My honest first read on Sol: "It's nowhere near Fable. Nowhere near Opus 5.5, so a little disappointing." The benchmark table below is why. They roll out in ChatGPT Work and Codex on paid plans. Neither one is in the Chat tab.
And on my own $100 Pro account, the afternoon they launched, they were not in Work yet either. My desktop app had an update downloaded but not installed, so that list came from the older build. I will show you.
Astra is the top rung, Sol is the middle, Luna is the bottom
Here is the whole family on one table, using OpenAI's own words for each model and its API prices.Rung
OpenAI's own line
API price per 1M tokens (in / out)
Leave it on forGPT-6 Astra
"Our most capable model, built for the hardest end-to-end work"
$10 / $50
Hard jobs: long documents, messy research, a real sparring partnerGPT-6 Sol
"Built to power complex coding and agentic workflows"
$2 / $10
Everyday workGPT-6 Luna
"high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions"
$0.10 / $0.50
Grunt work you repeat a lotPrices and the first two lines are from OpenAI's model docs, read September 22. Luna's line is OpenAI's wording as quoted in TechCrunch's launch report.
Read the price column as a ladder too. Astra costs 5 times what Sol costs per token. Sol costs 20 times what Luna costs.
That's the ladder I'd use: Astra for the hard jobs, Sol for the everyday work, Luna for the grunt work.
If the names still feel like a riddle, you are in good company. On Hacker News the day it launched, dyauspitr asked:"Wtf is GPT-6 Sol, I though GPT-6 is Astra?"Readerium answered with the best two-line explanation I have seen:"Number is generationName is the size (Luna smallest to Astra largest)"That is the system. The thread is worth a scroll if you want to see how many people are confused.
One more thing that trips people up: there is no GPT-6 Terra. The GPT-5.6 family was Sol, Terra and Luna. GPT-6 swapped Terra out and put Astra on top.
Sol and Luna live in Work and Codex, not in the Chat tab
OpenAI's announcement, reposted in its community forum, says it plainly: "GPT-6 Sol and Luna roll out today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Both are also available in the API." Launch coverage from 9to5Mac adds that Free and Go users get Luna in the desktop app.
So the Chat tab is not where you will find them. If Work is new to you, what ChatGPT Work is covers it in plain English.
For which plan gets what on which surface, I keep one table on which ChatGPT plans get Astra, and I am not going to duplicate it here.
Now the part the announcement cannot tell you. This is the Work model list on my Pro account on September 22, the afternoon of the launch:My ChatGPT Pro Work tab, September 22, 2026, pre-update: Astra, but no GPT-6 Sol or Luna yet.
And here is the same list twelve days earlier, for comparison:Same account, same tab, September 10, 2026. No Astra, and the names did not even carry the "GPT-" prefix yet.
Two things to take from that. "Rolls out today" and "is in your list today" are different days. And when GPT-6 Sol does land, it is going to sit right next to GPT-5.6 Sol, so check the number, not the name.
If yours is missing too, update the desktop app before you conclude anything about your plan. Mine had an update downloaded but not yet installed when I took that screenshot, so a newer build may already show them. That is the same step OpenAI documented when GPT-6 Astra went missing for people.
Leave Plus on Sol for everyday work and save Astra for the hard jobs
Would I tell a Plus user to leave Work on Sol for the everyday stuff and save Astra for the hard jobs?
Yeah, I would.
Here is the honest part. I got used to using Astra, not gonna lie. But I run it on low, because it burns usage fast.
That is literally a complaint, and it's legit. It's not efficient right now. It burns tokens.
The numbers back me up. OpenAI's own allowance table from September 10 puts Astra at 5 to 45 messages per five hours on Plus, against 10 to 100 for GPT-5.6 Sol. That is roughly double the usage for the same conversation, and I walk through every tier in Astra's usage limits.
That table is for the old Sol. OpenAI has not published ChatGPT limits for GPT-6 Sol or Luna yet.
The same OpenAI help article is also why running Astra on low is not crazy. It says Astra at Low effort can outperform Sol at High effort. That was written about GPT-5.6 Sol, so treat it as a hint, not a rule for the new one.
So my advice is simple: Sol stays selected for the everyday stuff. Astra comes out for the hard job, on low effort first.
Sol's real wins are price and a reliability claim
Sol deserves a fair hearing, and it has two genuine wins.
The first is price. At $2 in and $10 out per million tokens, Sol is one fifth of Astra's API price. An OpenAI spokesperson told VentureBeat these "are permanent prices, not promotional or introductory pricing."
The second is reliability. OpenAI says, per TechCrunch, that "GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost." That is OpenAI's own internal test, not an independent one.
OpenAI's Tibo (Thibault Sottiaux) went further, in a post the community thread reposts: "GPT-6 Sol and Luna are out. Not only are they a very significant improvement across the board, but also in writing and general 'you know when you try it' quality."
My read, before I have run it: Sol should be good for a lot of the stuff. I expect it to be better than GPT-5.6 Sol, and more efficient. OpenAI's claim points the same way.
OpenAI compared Sol to Opus 5 at medium effort, and Opus 5.5 shipped the same day
Here is where I am a little disappointed.
My opinion, going by the benchmarks and not my own runs: OpenAI basically dropped a cheaper, lower-tier model. It's barely at Opus 5 level.
That read holds up on OpenAI's own chart, with one caveat you should know about:Model
Effort setting
OSWorld 2.0 (computer use)
Cost per taskGPT-6 Sol
xhigh
60.5%
About 80% lowerClaude Opus 5
medium
60.3%
BaselineThese are OpenAI's numbers, as reported by VentureBeat on September 22.
Look at the effort column. Sol's effort dial runs none, low, medium, high, xhigh, max, so xhigh is its second-highest setting. Opus 5 was at medium.
Sol had to run near its top setting to tie Opus 5 at medium, and by OpenAI's math it still did that for about a fifth of the cost.
Then Anthropic released Claude Opus 5.5 hours before OpenAI's announcement. Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work, at $4 in and $20 out per million tokens. OpenAI's material has no Opus 5.5 comparison.
In fairness to Sol, it is still half of Opus 5.5's price per token. I compare Opus 5.5 with Fable properly in Opus 5.5 vs Fable.
The way I see it, 5.5 is a huge leap. You literally got a better model and it's cheaper.
Sol is a sidekick for grunt work, not my main model
Does Sol feel like a real option to me? To be honest, no.
I think it's gonna be a good sidekick for research and the grunt work I just want to burn usage on. If I want to tinker around and brainstorm, I'm still gonna use Astra.
So I have to find a use case for it. That is the honest status on day one: I have not run GPT-6 Sol or Luna on a real job yet, and it was not even in my list when I wrote this, on an app build one update behind.
I also have a complaint about the model count, and then I have to take half of it back.
I think this whole game of having five different models makes it seem like you have more options, when really you've got one or two good things. That's what I like Anthropic for right now, because it's not complicated. You've got Haiku, Sonnet, Opus and Fable.
So they do have a lot of models too, actually. Never mind.
But I rarely use Haiku. I rarely use Sonnet. I'm just using Opus and Fable. If that is where you are too, how I run Claude is the setup I actually use.
What nobody knows yet
A few things were not public when I wrote this, and I would rather say so than guess:ChatGPT limits for GPT-6 Sol and Luna. OpenAI has published no message numbers for either one.
When the Chat tab gets them. The announcement only names Work and Codex.
How much Luna a Free account gets. Launch coverage says Free and Go users get Luna in the desktop app, with no number.
Whether GPT-6 Sol still loses to Astra on Low. OpenAI's effort advice was written for GPT-5.6 Sol.I log plan and limit changes with dates on the AI plan tracker, and the box at the bottom of this page emails you when one moves.
Which one to pick depends on what you pay forIf you are on Plus: leave Work on Sol for everyday tasks once it shows up. Pull out Astra for the hard job, on low effort first, because it cost roughly double the usage of the old Sol.
If you are on Pro: you have room for Astra, and I still run it on low. Send Sol the bulk research you would rather not spend Astra on.
If you are on Free or Go: Luna in the desktop app is your only GPT-6 door. Use it for summaries, pulling details out of a document, and quick questions, which is exactly what OpenAI built it for.
If you also pay for Claude: Opus 5.5 is the one to beat now, and OpenAI did not even compare Sol to it. Treat Sol as a cheap sidekick, not a switch.Frequently asked questions
What is GPT-6 Sol?
GPT-6 Sol is OpenAI's middle GPT-6 model, released September 22, 2026. It sits under GPT-6 Astra and above GPT-6 Luna.
OpenAI's developer docs describe it as built to power complex coding and agentic workflows, and it costs $2 per million input tokens and $10 per million output tokens in the API, one fifth of Astra's price.
What is the difference between GPT-6 Sol and Luna?
Sol is the everyday model for real work. Luna is the small, cheap one for high-volume tasks with a clear goal, which OpenAI lists as summarizing documents, extracting information, or answering quick questions.
In the API, Sol costs $2 in and $10 out per million tokens, and Luna costs $0.10 in and $0.50 out, so Sol is 20 times the price of Luna.
Is GPT-6 Sol better than GPT-6 Astra?
No. Astra is the top of the GPT-6 ladder and Sol is the middle rung.
OpenAI says Sol reaches Astra-level reliability on its internal factuality test at much lower cost, but Astra is still the model OpenAI builds for the hardest end-to-end work. Use Sol for everyday tasks and save Astra for the hard jobs.
Where can I use GPT-6 Sol and Luna in ChatGPT?
OpenAI says both roll out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, and not in the Chat tab. Launch coverage (9to5Mac, VentureBeat) adds that Free and Go users get GPT-6 Luna in the ChatGPT desktop app.
On launch day afternoon, my own Pro account's Work list still showed GPT-5.6 Sol and Luna, on a desktop app with an update downloaded but not yet installed, so update the app first if yours is missing them.
Is GPT-6 Luna free?
Partly. Launch coverage from 9to5Mac and VentureBeat says Free and Go users get GPT-6 Luna in the ChatGPT desktop app, and OpenAI's help center had not caught up on launch day. Sol is for paid plans only. OpenAI has not published how many Luna messages a free account gets.
What happened to GPT-6 Terra?
There is no GPT-6 Terra. The GPT-5.6 family had three sizes, Sol, Terra, and Luna. The GPT-6 family announced so far has three different names: Astra on top, Sol in the middle, and Luna at the bottom.
Where this comes from
The facts come from OpenAI's docs and announcement, plus TechCrunch and VentureBeat, read September 22, 2026. The screenshots are my own Pro account, and the opinions are mine.
I have not run GPT-6 Sol or Luna yet. My list still showed the 5.6 versions, with an app update downloaded but not installed.
Published September 22, 2026. Last reviewed September 22, 2026.
Quick answer: GPT-6 Astra is OpenAI's most capable model, released September 3, 2026, with a context window over one million tokens. It is a model, not a new app.
And there is a good chance the word "Astra" does not appear anywhere in your ChatGPT. Your account is fine. It comes down to how OpenAI named things.
I pay $100 a month for ChatGPT Pro. When I opened the model list today to write this page, Astra was not in it.
More on that in a minute, because it turns out to be the most useful thing on this page.
What it is, without the benchmark table
Astra is the model OpenAI now points at its hardest work. From OpenAI's launch post: "state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work," with a context window of more than one million tokens.
OpenAI's own help center is blunter about when to reach for it. It calls Astra "Our most capable model for coding, research, analysis, and complex problem-solving," and the examples it gives are investigating a difficult bug and working through an unfamiliar problem.
Put that list in normal-person words. It does not say "answers your questions better." It says: reads a lot, goes and does things, sticks with a hard problem.
So in one line, Astra is for the jobs that take a while.
For me it is research and a sparring partner. It is good at pushing back on Claude Fable when I want a second opinion on something I already believe. And it is smart enough that the pushback is worth reading.
I have also started using it to build a game, which I will get to.
What I do not open it for is my daily content work. That still goes to Fable, because that is the one that feels like a partner rather than a very good search.
One model, three names
This is what turns a five-minute question into a forty-minute one, and nobody explains it.Where you are
What it is called thereChatGPT Work, and Codex (the developer surface, ignore it if you do not code)
GPT-6 AstraThe ChatGPT Chat tab
GPT-6 ProThe API
gpt-6-astraOpenAI's help center states the split directly: "In Chat, GPT-6 Pro is powered by Astra and is available on eligible Pro, Business and Enterprise plans. Its message limits are separate from the shared Work and Codex allowance and vary by plan. Plus includes Astra in Work and Codex, but not GPT-6 Pro in Chat."
So if you are on Plus, hunting the Chat model picker for the word "Astra," you will hunt forever. It was never going to be there.
Here is my own Chat picker on the $100 Pro plan a few days ago. No Astra anywhere in it.My own Chat picker, captured September 5, 2026, on ChatGPT Pro at $100 a month. The model is in the account. The name is not in the list.
Why you cannot find it, in the order worth checking
Now the live one. This is my Work tab today, September 10, with the model list open.My Work model list on September 10, 2026, on a Pro plan that is entitled to Astra. Default, 5.6 Sol, 5.6 Terra, 5.6 Luna, 5.5, 5.3 Codex Spark. No Astra.
I am not showing you that to claim OpenAI pulled the model. I am showing it because it is the most common thing that happens to real people in week two. OpenAI even documents the fix in its own help center, under a heading called "Update your ChatGPT desktop app."
Their instruction, which is more specific than the usual "try restarting":If Astra is missing, check for updates even if you recently installed one. Your app may need another update and a full restart.OpenAI's help center, captured September 10, 2026. A three-step procedure that exists because this happens to a lot of people.
So here's the checklist, in the order that actually fixes it:Update the desktop app, then fully quit and reopen. Then check for updates again. OpenAI says you may need a second round.
Check which tab you are in. Work and Codex offer Astra by name. Chat offers GPT-6 Pro instead.
Check your plan against the surface. Plus gets Astra in Work and Codex only.My app had crashed to an error screen minutes before I took that shot. When it came back, Astra still was not listed. So updating is a real step, not a brush-off.The state my app was in shortly before the model list above, September 10, 2026. Note which button OpenAI puts first.
When it is there, it looks like this. Same account, five days earlier, running a real job in Work:The same account on September 5, with the Work composer chip reading GPT-6 Astra Ultra. Third name, same model.
Do you have it? The short versionFree: no.
Plus: yes, in Work and Codex. Not in the Chat picker.
Pro $100 and Pro $200: yes, in Work and Codex, plus GPT-6 Pro in Chat.
Business and Enterprise: yes, with seat-level rules.I am deliberately not rebuilding the full surface-by-plan table here, because it already exists: which ChatGPT plans get Astra is the page for that question. This one is about what the thing is.
The part nobody warns you about: it eats your plan
Astra burns your included usage fast. I am on the $100 plan and I have already burned through two full resets.
That is not a vibe. OpenAI publishes estimated messages per five-hour window, and on every plan tier Astra's range is about half of GPT-5.6 Sol's. Same messages, at least twice the burn. The full table and the two clocks behind it are on GPT-6 Astra usage limits explained.
The number worth carrying around: on Plus, the bottom of OpenAI's own estimated range for Astra is five messages in a five-hour window. So think "a couple of messages" for a heavy Astra session.
That's why the setting nobody touches matters more than the model you pick. OpenAI's guidance here surprised me:Lower effort does not mean lower capability across models. For example, Astra at Low effort can outperform Sol at High effort. If you've been getting good results with Sol at High, try Astra at Low or Medium as a starting point.OpenAI's help center, captured September 10, 2026. The most useful paragraph OpenAI published about Astra, and it is buried under a settings heading.
I had heard that claim secondhand and half believed it. It is real, and it is in their documentation.
My own rule ends up in the same place, just from the other side. In my words: use the top setting to plan the hard thing, then drop down to execute it. How far down? Somewhere between low and high. I do not know.
Planning is where the extra reasoning is worth paying for. Doing the work mostly is not. Matching my "top setting" to OpenAI's effort control is my guess, not their instruction.
The full mechanics of the two clocks and three pools are in GPT-6 Astra usage limits explained.
What people are actually building with it
The word-of-mouth is people posting what they made. These are the ones that made me open the model picker. I linked the original posts so you can judge the source yourself.Tom Krcha gave Astra an old drawing of a steam train and asked for it in Blender. His post says it came back as 3,295 editable objects in a few minutes. Original post on X.Roberto Nickson fed it nine phone photos of his studio. It worked out the spatial layout and returned a walkable 3D model. Original post on X.
Three more worth your time, no screenshots because they are threads and video:KP's roundup of the early demos, compiled the day after launch with credit to each builder. The fastest way to see the range.
Min Choi's ten examples: games, 3D scenes, character models, all from prompts or images.
Riley Brown's take: his biggest moment was not visual at all. He gave Astra context on his whole business and let it plan. That matches my own use more than the game clips do.One caveat on all of it. Every one of these people is on a Pro or Business plan or the API. None of them are doing this on the $20 plan, and the next section is why.
The game thing, honestly
The hype about people building games with Astra is real. I have seen shooters. I wanted in, so I spent a weekend trying to build a Pokemon-style game with it.
The graphics it made stunned me. These looked different.Three screens Astra generated for me over the weekend of September 6. Original creatures, original names, a battle system and an overworld. This is what "impressive graphics" meant.
The game did not get built. It burned two full usage resets and what I ended up with was ugly.
I am fairly sure that is me, not the model. I did not ask the computer the right way, and the learning curve is a real curve. But I want that on the record next to the demo videos. The demo videos are all wins. The real week-two experience includes an expensive failure.
If you want to build something serious with it, my read is you want the $200 plan. At $100 you feel the ceiling, and I never felt it with GPT-5.6 Sol.
What I would do this week
If you are on Plus: open the Work tab, not Chat, and give it one hard thing you have been avoiding. A long document you never read, a mess you have not untangled. That is the shape of task Astra is for. Do not burn it on questions.
If you are on Pro: try the thing OpenAI suggests and drop your effort setting a notch. You may find you were paying double for reasoning you did not need.
And if you cannot find it at all, update the app twice before you conclude anything about your plan.
Astra is a real step up on hard, long, multi-step work. Just don't treat it like a better chatbot, or you'll spend a month's allowance on a Tuesday.
Working out which subscription this affects? Which ChatGPT plans get Astra has the surface table, GPT-6 Astra vs Claude Fable 5.1 has the benchmark read, and GPT-6 Astra vs Grok Bot covers the model-versus-teammate question. More of how I run all of this at Claude Guides.
Published September 10, 2026. The effort-level guidance, the Chat-versus-Work split and the update procedure were read and captured that day from OpenAI's help center article "Managing usage with GPT-6 Astra in Work and Codex," and the two screenshots of it above are from that capture. The allowance figures are summarized here and screenshotted in full on the limits post linked above. The launch description and context window come from OpenAI's September 3 announcement, cited as of September 4 when this site first captured it. The account screenshots are my own paid Pro plan, sidebar cropped. My missing-Astra screenshot documents one account on one day and one app build. It is not a claim about OpenAI's rollout.
Quick answer: these two models now cost exactly the same and are good at different things. GPT-6 Astra wins computer use, math, cybersecurity and long context. Claude Fable 5.1 wins the two broadest "how smart is it" boards, including one published by OpenAI itself.
Both are $10 per million input tokens and $50 per million output. So price isn't what decides it anymore.
Every headline yesterday said Astra swept. OpenAI's own launch table does not say that.
The rows nobody quoted
OpenAI published a comparison table on the Astra launch page with Claude Fable 5.1 in a column. Two lines in that table go to Claude.
Every number below is transcribed from that page, checked September 4, 2026. When OpenAI's footnotes put a caveat on a Claude score, I flag it in the next section.Benchmark (OpenAI's own table)
GPT-6 Astra
Claude Fable 5.1Artificial Analysis Intelligence Index v4.1.1
61.2
65.7Humanity's Last Exam (with tools)
57.2%
65.0%Terminal-Bench 4.0
57.9%
55.8%Terminal-Bench Science 0.1
64.6%
52.6%FrontierMath Tier 4 (v2)
97.6%
87.8%GPQA Diamond
96.0%
93.7%AutomationBench
41.4%
31.4%DeepSWE v1.1
74.1%
67.4%FrontierCode 1.1 Main
53.3%
50.9%ARC-AGI-2
95.0%
90.0%Computer use safety, internal (lower is better)
2.4%
9.5%The Artificial Analysis Intelligence Index is about as close as you get to one number for how smart a model is overall. OpenAI put it in its own launch table, and it hands Claude a 4.5 point lead.
Artificial Analysis, quoted independently:Sits beside GPT-5.6 Sol in Intelligence: GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback).So on the broadest board, OpenAI's newest model tied its last one.
Read the footnotes before you read the table
OpenAI ran these numbers. That does not make them wrong, but the footnotes change how you should read four rows.BenchCAD: OpenAI's footnote 5 says Claude's scores reflect three modifications to the eval.
ScreenSpot-Pro and ExploitGym: footnote 17 says the Fable scores reported are actually from Mythos, described as "Fable with fewer safeguards." That is a different model.
HealthBench Professional: footnote 11 says OpenAI independently evaluated the Claude models itself, using GPT-5.4 as grader, with Opus 5 substituted when Fable 5.1 refused.
Three science evals: footnote 12 says Claude Fable 5 and 5.1 are excluded from LifeSciBench, GeneBench Pro and MedChemBench because they refuse the majority of questions.None of that is shady. Every vendor benchmarks its own launch. But "state of the art across the board" is how the press summed it up, and the table doesn't say that.
The 99.9% asterisk
The number that traveled furthest was Astra saturating ARC-AGI-3 at 99.9%. It should count the least of anything in the launch.Two problems, both disclosed by the people who ran it:Harness. Per the ARC Prize blog, the 99.9% came from OpenAI's custom "Provider Adapter harness" at about $19K. The default ARC-AGI harness scored 62.7% at about $26K. OpenAI's own footnote 1 confirms it ran a modified responses API harness. That is a 37 point spread depending on plumbing.
No opponent. Claude Fable has no published ARC-AGI-3 result. So the biggest number of the launch has nothing to compare against.The Provider Adapter, in ARC Prize's words, "preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work." That's interesting engineering. But it doesn't tell you how smart the raw model is.
Simon Willison, writing the day it shipped, summed up the launch in one sentence: Astra "appears to score higher than Fable on most of OpenAI's self-reported benchmarks."
The key word there is self-reported.
Where Astra genuinely pulls ahead
Strip out the noise and Astra's wins are real. They just bunch up in a few areas.
Computer use. This is the big new thing in the launch. Astra scores 59.3% on Agents' Last Exam against 55.5% for Claude Opus 5, and OpenAI reports it hitting 72.6% on OSWorld 2.0 in roughly 47% less time per task than GPT-5.6 Sol.
Math and science. FrontierMath Tier 4 at 97.6% against 87.8%. Terminal-Bench Science at 64.6% against 52.6%. These are not close.
Cybersecurity. 100% on ExploitBench, 88.0% single-attempt on SRE-Bench binary reverse engineering. OpenAI says this crosses the Critical threshold in its own Preparedness Framework, so it put limits on the model at launch.
Long context. 100% on OpenAI's eight-needle test at 256K to 512K tokens, and 96.3% at 512K to 1M.
Token efficiency. This one matters more than the scores. On Agents' Last Exam, OpenAI reports Astra using about 65% fewer output tokens than Opus 5 at the highest-scoring settings.
Where Claude Fable 5.1 holds
The broad boards. Intelligence Index and Humanity's Last Exam, both from OpenAI's own table.
Coding, closer than the headline. Astra edges Fable 5.1 on FrontierCode 1.1 Main, 53.3% to 50.9%. But look one column over in OpenAI's table: Fable 5 scores 53.5%, ahead of Astra. On the Artificial Analysis Coding Agent Index, Astra posts 67.0 against 67.2 for Fable 5 and 68.1 for Opus 5. Coding is a wash.
Cost per task at equal quality. Artificial Analysis again: "Per task, the model is less than half the cost of Claude Fable 5, for the same score." That's where Astra's efficiency actually pays off. And it matters more than the trophy rows.
What I actually threw at it on day one
I pay for both stacks, and Astra is one day old as I write this.I didn't run a benchmark. I just ran my actual work, because that's the only test I trust.
The one that changed my mind about Astra was a prospecting task. I asked it to find companies in a specific niche worth reaching out to, and it went and searched forums and threads on its own without me telling it where to look.
What came back was about a month old. That kind of search usually pulls up threads from many months back.
For me that's the real upgrade: it did the digging on its own.Another Astra run from the same week, on my Pro account. I asked it to plan two weeks of paid-work experiments around my real numbers; it pulled live listings on its own and priced the options. The top of the answer has my personal figures in it, so it is trimmed. The chip in the composer is the part worth noticing: on this surface it is labeled GPT-6 Astra Ultra, not GPT-6 Pro.
Then I did the thing I would actually recommend: I ran the output back through Claude Fable to QA it, Fable rewrote the prompt, and I sent it back at Astra. The second pass looked good enough that it's now how I'd run this by default.
Now the other side. I've used Fable more, and for building it still feels more familiar and more dependable to me. But Fable 5.1 has also done some dumb things this week, mostly losing track of tools it definitely had access to. It's minor, but it's real.
My read on the scoreboard: I assumed it said Astra beat Fable outright. It doesn't, and that actually matches what I saw hands-on. Fable still feels better at the deep work. Astra feels newer at getting things done by itself.
The decision rule I would give someone today
Don't switch providers over this. Send each job to the one that's better at it.The job
OpenDeep building, writing in your voice, long careful work
Claude Fable 5.1Anything the AI should do on its own across apps and the web
GPT-6 AstraResearch and volume where you need lots of runs
GPT-6 AstraReviewing and QAing another model's output
Claude Fable 5.1Math, science, anything with a checkable right answer
GPT-6 AstraYou can only pay for one, and you are a generalist
ChatGPT (usage and resets decide it, not the score)I keep coming back to this: smartest isn't the flex anymore. A model that burns tokens to win a board is worth less to me than a leaner one that finishes the task. Both companies are getting pushed that way, and this launch is the clearest sign of it so far.
Verdict
Same price, split by job, and that split is the useful part. Astra is the better agent. Fable is still the better thinker on the broadest boards OpenAI itself published, and it's my daily driver for building.
If you were about to cancel Claude because of a headline, read the table first.
Working out which subscription this actually affects? Which ChatGPT plans get Astra, Claude Max vs ChatGPT Pro, and Claude vs ChatGPT for the everyday call. More of how I run both at Claude Guides.
Published September 4, 2026, one day after GPT-6 Astra shipped. Every benchmark figure was transcribed that day from OpenAI's launch page, and the ARC-AGI-3 harness figures from Simon Willison's write-up citing the ARC Prize blog, both linked above. These are vendor-run numbers on a vendor page and OpenAI's footnotes qualify several Claude scores; read the footnotes section before quoting any row. Benchmarks and prices move fast and this page will need re-checking.
Chris Alarcon 
