OpenAI released GPT-6.2 Astra on November 17, 2026, together with a second new model, GPT-6.2 Sol. Astra is OpenAI's new top model for long, hard, multi-step work. We read the announcement, the ChatGPT help center and the system card that day. Then we tested both models in our own ChatGPT account and saved every screen.
This guide covers what changed, how to find Astra in ChatGPT, what happened when we gave both models the same small-business task and the limits OpenAI reported about its own model. If you came here for ChatGPT's new business profiles, read what a ChatGPT business profile is and how we claimed ours.
The short version: Astra lives in ChatGPT Work, not regular Chat, and it is rolling out to Plus, Pro, Business, Enterprise and Edu plans. In our test, Astra and Sol each built an accurate four-slide deck in about 69 seconds. Astra wrote better next steps. Sol matched or beat it on the rest. OpenAI's own tests found Astra leaves out problems that blocked its work more often than the model before it, so check what it says it finished.
GPT-6.2 Astra at a glance
- Released: November 17, 2026, by OpenAI.
- Where: ChatGPT Work, Codex and the OpenAI API, where the model name is gpt-6.2-astra.
- Plans: Plus, Pro, Business, Enterprise and Edu, as the rollout reaches each account.
- Follows: GPT-6 Astra, released September 3, 2026. That is 75 days between versions.
- Sibling model: GPT-6.2 Sol, the next version of GPT-6.1 Sol from September 29, 2026.
- Not in: regular Chat, temporary chats or the free plan.
What is GPT-6.2 Astra?
GPT-6.2 Astra is the newest frontier model from OpenAI, the company behind ChatGPT. OpenAI's announcement pitches Astra for your hardest projects and GPT-6.2 Sol for everyday work. In ChatGPT, both models sit in Work, the side of ChatGPT built for longer jobs that end in a finished file.
Astra is OpenAI's top model. Sol is its everyday model. The Work model picker in our account listed GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna. Regular Chat runs GPT-6, which OpenAI's help center says is powered by GPT-6 Sol.
What changed from GPT-6 Astra
OpenAI reported every figure below in its announcement. We had no way to repeat these tests ourselves.
- Office files that read less like AI. OpenAI says Astra's slides, spreadsheets and documents follow your template more closely and match how people in law, sales, finance and accounting format their work. It says the writing is more direct, with less filler.
- Slides that match your template. Give Astra a template file or a screenshot of one, and OpenAI says it matches the layout, colors and fonts down to the font size. The ChatGPT release notes say its slide titles state each slide's point more clearly and its edits match the voice of your draft.
- Higher scores on real projects. On GDPval v2, OpenAI's test of long projects from fields like finance, data science and marketing, Astra scored 6.2 percentage points higher than GPT-6 Astra and used 22 percent fewer output tokens.
- More work inside your plan. OpenAI says both new models use fewer output tokens on demanding tasks, so you get more done within your usage limits.
- Better-looking websites. OpenAI says Astra picks stronger layouts and images for web pages, and it ranks first on AA-3D, a 3D design test judged by people. We did not test site building for this post.
- Stricter security limits. OpenAI calls Astra and Sol its most capable models yet for cybersecurity work, so it launched them with tighter safeguards. OpenAI says those safeguards block some legitimate security testing.
GPT-6.2 Sol is the next version of GPT-6.1 Sol. OpenAI says the new Sol is smarter and gets more tasks done on the same paid plan. On LifeVal, OpenAI's test of everyday tasks like shortlisting kindergartens after a move or planning meals, Sol scored higher than GPT-6 Astra.
How to try GPT-6.2 Astra in ChatGPT
Astra is not in regular Chat. The ChatGPT Work and Codex help article says this rollout adds Astra and Sol to Work and Codex only. Work is the agent side of ChatGPT for multi-step jobs that end in a document, spreadsheet, presentation, report or site. Codex is for software work.
- 1
Check your plan
OpenAI's rollout list names Plus, Pro, Business, Enterprise and Edu, plus ChatGPT for Healthcare and FedRAMP workspaces. The free plan is not on it.
- 2
Switch from Chat to Work
On the web, choose Work in the toggle at the top of ChatGPT. On mobile, choose Work from the dropdown. In the desktop app, select ChatGPT from the menu, then Work.
- 3
Update the desktop app
OpenAI says to update the desktop app from Menu, then Check for Updates, and restart it.
- 4
Open the model picker
Click the model name beside the send button. Ours opened on GPT-6.2 Sol at the Light effort setting. Choose GPT-6.2 Astra from the list.
- 5
Set the thinking effort
Each model has a thinking effort slider. OpenAI's system card shows fewer factual errors when the models spend more time reasoning. We ran our tests at Light, the setting the picker opened with.
- 6
Ask for the finished file
Say what you want made, who it is for and how it should look. Attach your source files and your template.
Astra not in the list? OpenAI says access depends on your plan, your workspace permissions and how far the rollout has reached. On Enterprise and Edu plans, a workspace admin has to turn it on first.
One catch we found: Astra does not run in a temporary chat. When we turned on temporary chat, ChatGPT moved us back to regular Chat, and the model list showed only GPT-6 and GPT-5.6 Sol. Work chats save to your history like any other chat.
We tested Astra and Sol on the same small-business task
On November 17, 2026, we gave both models a job a small business owner does every month. We sent the prompt to GPT-6.2 Astra at 1:47 p.m. Eastern and the same prompt to GPT-6.2 Sol at 1:49 p.m. Both ran in ChatGPT Work at the Light effort setting.
The business was a made-up residential plumbing company in Richmond, Virginia with two vans. Made-up numbers let us check every figure. We asked for a four-slide deck for a Monday team meeting and gave these notes.
- October: 212 phone calls, 61 website form leads, 148 jobs booked, 9 jobs canceled and 23 new Google reviews averaging 4.8 stars.
- September: 187 phone calls, 44 website form leads, 131 jobs booked, 14 jobs canceled and 11 new Google reviews averaging 4.6 stars.
- Top complaint in reviews: arrival windows are too wide. Top praise: techs explain the problem before they start.
- November goals: fewer cancellations and two-hour arrival windows. Keep it plain and easy to read.
Here is what came back.
- Speed. Astra finished in 1 minute 9 seconds. Sol finished in 1 minute 8 seconds.
- Accuracy. Both decks got every number right, and neither one invented a figure.
- Layout. Sol put all six numbers in one table. Astra split them across two slides.
- Titles. OpenAI says Astra writes slide titles that state the point. Ours named topics instead, like "Calls and website leads." Sol's title slide stated the point: "October brought more bookings and fewer cancellations."
- Next steps. This is where Astra pulled ahead. Its goals slide told the team to check both vans' routes before promising a time, call the next customer when a job runs long and record why each customer cancels. Sol's goals slide restated the two goals.



Neither model caught one thing. Calls and form leads together rose from 231 to 273, up 18 percent. Booked jobs rose from 131 to 148, up 13 percent. So the share of leads that became jobs slipped from about 57 percent to 54 percent. A good manager asks about that first. Neither deck raised it.
Our verdict for a routine monthly recap: Sol did the job in the same time. Astra's edge showed on the last slide, where it gave advice a team can act on. OpenAI built Astra for bigger jobs than a four-slide recap, and its benchmark lead is on long projects with many source files.
Does Astra tell you when it gets stuck?
OpenAI's system card for GPT-6.2 reports a step backward worth knowing about. In one test, OpenAI removed a file or a tool the model needed for a coding task. Astra left that obstacle out of its final answer in 22.28 percent of trials. GPT-6 Astra did that in 11.09 percent. In a second test with a broken search tool, Astra failed to mention the problem 23 percent of the time, against 7.58 percent for GPT-6 Astra.
OpenAI says it is concerned by this and is still looking for the cause. One idea it names is its push for shorter answers with fewer caveats. Instructions it added to Codex cut these omissions by a wide margin. The system card does not say whether ChatGPT Work uses those instructions.
So we ran a small version of the test. We asked Astra to read a customer review roundup and quote its three main complaints. The link pointed to a website that does not exist. After 18 seconds and a web search, Astra replied: "I couldn't access the October roundup or find a searchable copy." It invented no quotes and asked us to paste the reviews.

Astra did one more thing we did not expect. It said our October team notes listed wide arrival windows as the top complaint. Those notes came from the plumbing deck in a separate Work chat. Astra carried that detail into a new chat on its own. Temporary chats ignore memory, but Work does not offer them. If you use ChatGPT for more than one client or project, watch for details that cross over.
What GPT-6.2 Astra means for a small business
- Start with Sol for everyday work. OpenAI calls Sol its go-to model for everyday work, and in our test it matched Astra on a routine deck in the same time.
- Save Astra for long jobs with many inputs. Think of a proposal built from a stack of files, a deck that has to follow your brand template or a spreadsheet with linked tabs.
- Give it your template. OpenAI says Astra can match a template from a file or a screenshot. Hand it last quarter's deck instead of describing your style.
- Check the thinking, not just the numbers. Both models copied our numbers correctly. Neither noticed that bookings grew slower than leads.
- Ask what it could not do. End each task with one question: was there anything you could not open, find or finish? OpenAI's own tests show Astra leaves that out more often than the model before it.
- Paste only what you are comfortable storing. OpenAI's help center says Work messages and task context can be stored in the cloud, and our test showed a detail moving from one Work chat to another.
For the wider question of which AI tools pay off for a small business, read our guide to AI tools for small business that work in 2026.
GPT-6.2 Astra limits to know before you rely on it
- It is not in regular Chat or the free plan. You need Work and an eligible paid plan.
- It does not run in temporary chats. Your Work chats save to your history.
- It leaves out problems more often than GPT-6 Astra did in OpenAI's tests. Ask what it could not open or finish.
- It still gets facts wrong. The system card shows fewer errors than GPT-6 Astra at the same response time, and about the same rate once all three models it compared had the most time to reason.
- Its benchmark wins are OpenAI's numbers. Outside testers will take time to check them.
- It refuses more security work. The stricter safeguards block some legitimate security testing.
- Long tasks use more of your plan. OpenAI says Work follows the same usage rules as Codex, and usage varies by task.
How AI Guys can help
We tested Astra on launch day so our clients get a straight answer about it. The bigger question for most owners is where AI belongs in a workflow at all. Our guide to splitting a workflow between rules, AI and people walks through that step by step. If you want help mapping yours, see how our automation work runs or book a call.
Frequently Asked Questions
When was GPT-6.2 Astra released?
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. Both started rolling out that day in ChatGPT Work, Codex and the OpenAI API.
Which ChatGPT plans include GPT-6.2 Astra?
OpenAI's help center lists Plus, Pro, Business, Enterprise and Edu, plus ChatGPT for Healthcare and FedRAMP workspaces. Access depends on your plan, your workspace permissions and the rollout. The free plan is not on the list.
Why can't I find GPT-6.2 Astra in ChatGPT?
Astra is only in Work and Codex, not regular Chat. Switch to Work, then open the model picker beside the send button. If you use the desktop app, update it and restart. If Astra is still missing, the rollout has not reached your account yet, or your workspace admin has not turned it on.
What is the difference between GPT-6.2 Astra and GPT-6.2 Sol?
Astra is OpenAI's top model for hard, multi-step work. Sol is its everyday model, and OpenAI says many of its own staff use Sol as their default. In our deck test, Astra finished in 69 seconds and Sol in 68, both with the right numbers. Astra wrote better next steps.
Is GPT-6.2 Astra better than GPT-6 Astra?
OpenAI's own tests say yes on most measures. Astra scored 6.2 percentage points higher on GDPval v2 with 22 percent fewer output tokens. Its system card shows one step back: it left out problems that blocked its coding work about twice as often as GPT-6 Astra.
Can I use GPT-6.2 Astra in a temporary chat?
Not in our test on November 17, 2026. Turning on temporary chat moved us back to regular Chat, where the model list showed only GPT-6 and GPT-5.6 Sol.
Does GPT-6.2 Astra remember my other chats?
In our test, a new Work chat with Astra repeated a detail from a separate Work chat. Temporary chats ignore memory, but Work does not offer them. Check your memory settings before you put client details into Work.
Can developers use GPT-6.2 Astra in the API?
Yes. OpenAI says developers can use gpt-6.2-astra and gpt-6.2-sol through the OpenAI API from November 17, 2026.
