GPT-6.2 Astra Is Cheaper. Is It Better for Your Work?
OpenAI released two new models today and cut the price of its flagship by a fifth. Its charts say the new one does more for a third of the cost. The first independent test we have seen agrees on the cost and the speed, and found it got slightly fewer answers right than the model it replaces.
What OpenAI released
On 17 November OpenAI released two models. GPT-6.2 Astra is described in its announcement as the new frontier model for your hardest projects. GPT-6.2 Sol is described as the efficient, versatile model for everyday work.
Both are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Enterprise and Edu users get access once their workspace administrator enables it. Developers can use both through the API.
| GPT-6.2 Astra | GPT-6.2 Sol | |
|---|---|---|
| Positioned for | Hardest projects | Everyday work |
| API input, per million tokens | $8 | $2, as listed by price trackers |
| API output, per million tokens | $40 | $10, as listed by price trackers |
| Cached input | $0.40 | $0.10, as listed by price trackers |
| Context and output limits | About 1.05 million tokens in, 128,000 out | The same, as reported |
Prices are in US dollars. OpenAI's announcement gives the Astra figures directly and says they are 20% below GPT-6 Astra on input and output, with a 60% cut on cached input. The Sol figures come from third-party price listings on launch day.
One detail for anyone comparing vendors: $8 and $40 is exactly what Anthropic charges for its top general model, as covered in our look at Claude Fable 5.2. The two flagship models now cost the same per token.
What OpenAI claims
The headline claim is about cost per task, not per token. OpenAI says GPT-6.2 Astra costs significantly less per task than GPT-6 Astra because it uses fewer tokens to finish the same work, on top of the lower prices.
Its own charts show the following. All of these are the vendor's figures.
| Benchmark, as described by OpenAI | Claim |
|---|---|
| Agents' Last Exam, multistep professional work | 5.2 points above GPT-6 Astra at less than a third of the cost |
| GDPval v2, an internal benchmark | 6.2 points above GPT-6 Astra, using 22% fewer output tokens |
| Terminal-Bench-Science | 9 points above GPT-6 Astra at less than a third of the cost |
The announcement also compares Astra favourably with a competitor's model on one of these tests. A vendor's chart of a rival is the least reliable kind of benchmark, because the vendor chose the test, the settings and the effort level for both sides. We have left that number out.
What one independent test found
Plotly ran the new model through its own data analysis exam on launch day, and the result is more useful than any leaderboard because it shows both sides of the trade.
| On Plotly's 43-question exam | GPT-6.2 Astra | GPT-6 Astra |
|---|---|---|
| Questions passed | 41 | 43 |
| Cost of the full run | $4.11 | $19.97 |
| Median time per question | 12.4 seconds | 26.9 seconds |
So on that test the new model was about a fifth of the cost and about twice as fast, and it got two fewer questions right than the model it replaces. Plotly notes that it tied six other models at 41 and that eight of the 53 models on its board scored higher.
That is one exam of 43 questions on one kind of work, so it proves little by itself. It is a useful corrective to the word frontier, though. Cheaper and faster is a real improvement. It is a different improvement from more accurate, and for some jobs you would rather have the two extra right answers.
What it means if you use ChatGPT
If you are on a Plus, Pro, Enterprise or Edu plan, you have the new models in ChatGPT Work and Codex without doing anything, subject to your administrator on the organisation plans. The announcement as we read it does not mention the free tier, and it does not change subscription prices, which are in our guide to ChatGPT pricing in Canada.
The practical step is small. Take three or four pieces of real work you did with the previous model, such as a report summary, a spreadsheet analysis and a customer reply, and run them again. Compare the results yourself. Your own tasks are a better guide than a benchmark built for somebody else's, which is the argument in how benchmarks mislead.
Use Sol for routine work and keep Astra for the difficult jobs. That is how OpenAI positions them, and it matches the advice Anthropic gives about its own lineup: start on the cheaper model and move up only when the results say you need to.
What it means if you build on the API
A 20% price cut is welcome, and it is not a reason to switch models unattended. Re-run your evaluation set before changing anything in production. If you do not have one, this release is the prompt to build it: twenty or so real tasks with checked answers and clear scoring.
Measure cost per completed task, not cost per token. The whole of OpenAI's pitch rests on the new model using fewer tokens, and whether that holds on your workload is something only your own numbers will show. And remember what a quality test can and cannot tell you: it flags a change in output, it does not confirm which model served you, as set out in knowing which model actually answered.
If your product names a model in a contract or a customer promise, check whether this release changes what you are delivering. And keep the uncertain-answer path in place whatever the model, for the reasons in why a model that knows when it is unsure is worth more.
Frequently Asked Questions
What are GPT-6.2 Astra and GPT-6.2 Sol?
Two models OpenAI released on 17 November 2026. OpenAI describes Astra as its new frontier model for the hardest projects and Sol as its efficient, versatile model for everyday work. Both are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the API.
How much does GPT-6.2 Astra cost?
In the API, 8 dollars per million input tokens, 0.40 dollars per million cached input tokens and 40 dollars per million output tokens, in US dollars. OpenAI says that is 20 percent below GPT-6 Astra on input and output and 60 percent below on cached input. ChatGPT subscription prices are a separate matter and the announcement does not change them.
Is GPT-6.2 Astra better than the model it replaces?
On the benchmarks OpenAI published, yes, and those are the vendor figures. On one independent data analysis exam run by Plotly, it passed 41 of 43 questions against 43 for GPT-6 Astra, at about a fifth of the cost and twice the speed. Whether it is better for you depends on whether your work values the saving or the last few correct answers.
Do I need to do anything if I use ChatGPT?
No action is needed to get access on a paid plan, though Enterprise and Edu users need their administrator to enable it. It is sensible to rerun a few pieces of your own recent work on the new model and compare the results before relying on it for anything important.
Should we switch our application to the new model?
Test before switching. Rerun your own evaluation tasks, measure the cost of each completed task and not only the token price, and check anything in your contracts or customer promises that names a specific model. A price cut is welcome and is not evidence about quality on your workload.
How does the price compare with other vendors?
At 8 and 40 dollars per million tokens, GPT-6.2 Astra is priced the same as Claude Fable 5.2, the top general model from Anthropic. Token price is only part of the comparison, since models use different numbers of tokens to finish the same task.
Pick on your own tasks
We build a small test set from your real work and run it across models, so the choice rests on your results and your costs.
Related Articles
Fable 5.2 Is Out. Start on the Cheaper Model Anyway
Offline AI: When Your Tools Need to Work Without Wi-Fi
Small AI Models Are Getting Good: Bigger Isn’t Always Better
Ajan leads the ChatGPT.ca team: 200+ custom GPT builds and automation projects for 50+ businesses across 20+ industries. Based in Markham, Ontario. PIPEDA-compliant solutions.