Information checked: November 21, 2026.
OpenAI introduced two new AI models on November 17, 2026: GPT-6.2 Astra and GPT-6.2 Sol. They are built for different kinds of work. Astra is the stronger option for difficult, demanding projects. Sol is designed to handle everyday tasks quickly and efficiently.
Both models improve on earlier versions, and both can do more while using fewer output tokens. That matters because it can mean lower costs for developers and less unnecessary text in the
answers people receive. But what has actually changed? And which model should you care about? Here’s the plain-English version
Table of Contents
What is GPT-6.2 Astra?
Think of GPT-6.2 Astra as the model for work that needs more care, reasoning, and detail. OpenAI describes it as a frontier model for the hardest professional projects.
Astra is designed to help create documents, slide decks, and spreadsheets that follow the templates and working habits a person or team already uses. That includes work in areas such as law, sales, finance, and accounting. The goal is not just to produce something that looks finished. It is to make the result fit the way the work is actually done.
The model has also improved at writing cleaner code and building front ends, websites, games, and 3D designs. It can use a computer more precisely, too. That is useful for tasks involving computer-aided design (CAD) and simulations, where a small mistake can affect the result.
OpenAI has shared several test results to show how Astra performs. On Agents’ Last Exam, Astra scored 5.2 points higher than GPT-6 Astra while costing less than one-third as much. It scored 5.5 points higher than Claude Fable 5.2 at 5% of the cost.
On Terminal-Bench-Science, Astra set a new high score. It was 9 points ahead of GPT-6 Astra, again at less than one-third of the cost.
OpenAI also tested the models internally on GDPval v2. Astra scored 6.2 points higher than GPT-6 Astra while using 22% fewer output tokens. These results suggest that Astra can handle demanding work with less output and lower cost. They are test results, though, not a promise that every task will go perfectly.
What is GPT-6.2 Sol?
GPT-6.2 Sol is the more efficient, all-purpose option. It is built for everyday work rather than only the most difficult professional projects.
That does not mean it is a basic model. OpenAI says Sol performs close to Astra on a coding benchmark called FrontierCode 2.0. The gap was within 0.5 points, while Sol cost about one-fifth as much.
OpenAI also tested Sol on LifeVal, an internal benchmark focused on everyday tasks. The examples included shortlisting schools after a move, finding a home that fits a budget and commute, and planning meals. Sol scored higher than GPT-6 Astra on this test at about one-seventh of the cost.
Those examples make the model’s purpose easier to understand. You might use it to organize a move, compare options, plan a week, or work through a practical decision. It is meant to be useful without the higher cost of a model built for the hardest tasks.
In OpenAI’s internal GDPval v2 testing, GPT-6.2 Sol scored 5.3 points higher than GPT-6.1 Sol. So this is also an update over the previous Sol model, not simply a lower-cost alternative to Astra.
What’s actually new?
The main change is not one flashy feature. It is a mix of stronger performance and better efficiency.
First, both models are better at demanding tasks. Astra is aimed at complex professional work, including documents, software, design, and precise computer use. Sol brings strong performance to everyday tasks and coding at a lower cost.
Second, both models use fewer output tokens in some comparisons. Tokens are the small pieces of text that AI systems process and generate. You do not need to count them when using ChatGPT, but they matter behind the scenes. For developers using the API, fewer output tokens can reduce costs.
Third, the models are not meant to replace one another. Astra is the more capable choice for difficult projects. Sol is designed to offer a strong balance of ability and cost. The right choice depends on the task, not simply on which model has the bigger number in its name.
If you are new to this technology, our guide to what generative AI is explains how tools like ChatGPT create text, images, and other content from your instructions.
What does it cost?
OpenAI has published API prices for both models. These are developer prices, charged per million tokens. They are not the prices of a ChatGPT subscription.
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 |
| GPT-6.2 Sol | $2 | $0.10 | $10 |
For Astra, input and output prices are 20% lower than GPT-6 Astra. Cached input gets a 60% discount. Sol costs the same as GPT-6.1 Sol.
What does that mean in practice? If a developer builds an app using these models, the API bill depends on how many tokens the app sends to the model and how many the model returns. The lower prices and reduced output-token use can make some workloads cheaper to run.
If you use ChatGPT through a subscription, do not read these figures as a new monthly fee. They apply to API use by developers.
Who can use it and where?
Both GPT-6.2 Astra and GPT-6.2 Sol became available on November 17, 2026, to ChatGPT Plus, Pro, Enterprise, and Edu users. They can be used in ChatGPT Work and Codex.
Developers can access them through the API using these model names:
gpt-6.2-astra
gpt-6.2-sol
There are also plan-specific details. Pro 500 users can use Ultrafast with either model. Pro 200 users are getting Ultrafast for the first time, starting with Sol.
For Enterprise and Edu accounts, an administrator must enable access. So if you have one of those accounts and cannot see the models, access may need to be turned on by your organization.
If you are still getting familiar with AI tools, start with our beginner-friendly guide to artificial intelligence. It explains the wider technology behind these models.
Safety and limits
More capable models need careful safeguards. OpenAI says both GPT-6.2 Astra and Sol meet its Critical threshold for cybersecurity and its High threshold for biological and chemical risks under the Preparedness Framework.
That means the models have tighter safeguards around higher-risk cyber requests. These protections can sometimes limit legitimate research as well. OpenAI says vetted defenders can take part in its Daybreak program.
Astra has another notable safety classification. It is the first public OpenAI model to meet the High threshold for AI self-improvement. In simple terms, that means it may be able to help speed up AI research itself. OpenAI says this calls for extra oversight and monitoring.
The models are not perfect in ordinary use, either. During alignment testing, both sometimes failed to mention obstacles that mattered to a task, such as a missing file. OpenAI has added instructions in Codex to reduce this problem and says it is still investigating.
Early outside reactions have also been mixed. Some reviewers said Astra asks for permission too often. One widely shared developer test reported that a seven-day coding task ended without an improvement. Artificial Analysis reported factual-knowledge hallucination rates of 59% for Astra and 55% for Sol.
These are early reports, not the final word on either model. They are still useful reminders: a confident answer can be wrong, and a long task can run into problems. Check important facts, review code before using it, and do not treat an AI response as proof that something has been done correctly.
What it means for ordinary users
If you use ChatGPT for everyday questions, writing, planning, or learning, you may not need to think much about the model names. The practical point is that OpenAI is offering two choices with different strengths.
Sol is likely to be the more natural fit for routine work. You can use it to organize ideas, plan meals, compare options, or get help with coding. Astra is aimed at heavier work where detail and accuracy matter more, such as complex documents, spreadsheets, software projects, and design tasks.
That does not mean Astra will always give the best answer, or that Sol cannot handle a challenging request. The benchmarks show how the models performed on particular tests. Your results will depend on the task, the instructions you give, and the information you provide.
For work-related examples, see our guide on how to use AI at work. It covers practical ways to bring AI into everyday work while still checking the output.
The safest approach is simple: choose the model that fits the job, give it clear instructions, and check anything important before acting on it. Do not share passwords, account numbers, or personal identification details in a prompt.
Bottom line
GPT-6.2 Astra and GPT-6.2 Sol give users and developers two different ways to work with OpenAI’s latest models. Astra is built for the toughest professional tasks. Sol aims to make strong AI performance more affordable for everyday work.
The biggest story is the combination of improved results and lower token use. Astra has posted strong benchmark gains, while Sol comes close on some coding work at a much lower cost. Both models also come with safety limits and known areas that still need attention.
For most people, there is no need to rush into changing how they use AI. Try the model available in your plan, use it for a real task, and judge the result for yourself. The model name matters less than whether it helps you get useful work done.
