# OpenAI Releases GPT-6.2 Astra and Sol. Which Should You Use?

> Source: <https://www.digitalapplied.com/blog/gpt-6-2-astra-sol-release-pricing-access>
> Published: 2026-11-17 00:00:00+00:00

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. The surprise is how close the cheaper model comes on one demanding coding test. At maximum effort, Sol finished within half a point of Astra on FrontierCode 2.0 at about a fifth of the cost, according to OpenAI. Astra makes its case elsewhere, particularly when the job ends in a workbook, a presentation or a piece of software that somebody else has to pick up and use. [OpenAI’s release announcement](https://openai.com/index/introducing-gpt-6-2-astra-sol/)

If you open ChatGPT Work or Codex to tackle a familiar job today, try Sol first. If you have a report that repeatedly comes back for corrections, give Astra the original files and the version your reviewer signed off. The difference worth looking for is a specific one. Can the reviewer follow the numbers, change an assumption and still trust the result?

1. 01Sol keeps the price of GPT-6.1 SolIts standard short-context API rates are $2 per million input tokens and $10 per million output tokens.
2. 02Astra gets a price cutStandard input and output rates fall 20% from GPT-6 Astra to $8 and $40 per million tokens. Cached input falls to $0.40.
3. 03Both models launch in ChatGPT Work, Codex and the APIThe launch announcement names Plus, Pro, Enterprise and Edu for the two user-facing products.
4. 04Developers get the same context capacity on both modelsRequests above 272,000 input tokens cross a separate billing threshold well before either model reaches its input limit.

## 01 — The earnings deckWhat changed in *the work itself*

One of OpenAI’s most revealing examples is a quarterly earnings presentation. The old and new Astra both recognized that tariff refunds had lifted the retailer’s earnings. The new model put sales growth of 5.3% in the headline and separated online growth from store growth. That changes what an executive sees first. Sales can inform next quarter’s plan. A refund that may never recur belongs in a different part of the conversation. [OpenAI’s side-by-side examples](https://openai.com/index/introducing-gpt-6-2-astra-sol/)

You can borrow that test even if your work has little to do with earnings. Give the model last month’s client deck, this month’s results and one paragraph about a change in how a metric was collected. Ask for the revised deck. Look at the headline before reading the notes. If the change in measurement accounts for the apparent improvement, a reviewer should be able to see that on the slide where the number appears. A handsome chart can make a bad comparison more persuasive.

## 02 — Code reviewWhy Sol deserves the first coding job

FrontierCode tests a question familiar to anyone who reviews pull requests. Would the maintainer accept this change? The people who maintain the repositories help define the tasks and their grading criteria. Correct behavior matters, along with the tests that establish it, the scope of the patch and how well it fits the surrounding code. A solution that misses a blocking requirement can earn zero. [Cognition’s scoring methodology](https://cognition.com/frontiercode)

That makes OpenAI’s half-point gap more interesting than a close finish on a coding quiz. Think of the common request to add a field to an export. The field must appear under the right heading, carry through when the value is empty and leave older exports readable. Those are conditions a reviewer can check quickly. Sol’s result makes it a sensible first attempt. An intermittent discrepancy spanning billing, exports and historical records is a different proposition. That is a reason to give Astra the same failing case and compare the diagnoses before choosing a patch.

## 03 — RepeatabilityThe spreadsheet result has a demanding denominator

OpenAI reports that Astra answered 62.5% of AA-AnalystAgent questions correctly on all five independent attempts. Artificial Analysis runs each of 80 questions five times. On that design, 62.5% corresponds to 50 questions answered correctly every time. The other 30 could include questions the model answered correctly four times. You cannot derive its average single-attempt accuracy from that headline. [OpenAI’s result](https://openai.com/index/introducing-gpt-6-2-astra-sol/) · [Artificial Analysis’s scoring method](https://artificialanalysis.ai/articles/aa-analyst-agent)

This distinction matters for a report that goes out every Monday. Suppose an analyst asks for acquisition cost by channel. A model can divide spend by conversions perfectly and still pick the wrong conversions column after a vendor changes its export. Keep the source files for a week when that happened. An informative trial asks the model to produce an editable workbook, trace the denominator to its source and explain any break in the series. A reviewer can then change the conversion count in one cell and see whether the report updates correctly.

## 04 — Design and planningA paper lamp and a kindergarten shortlist

OpenAI’s most tangible demonstration starts with a building and ends with a lamp. Its team took the scalloped shape of Chicago’s Marina City balconies, worked through a series of computer-aided designs with Astra and assembled a paper lamp by hand. The interesting step is the last one. A rendering can hide a seam or a tab that is impossible to reach. Once a person has to fold the paper, the design has to answer to the material. [OpenAI’s lamp demonstration](https://openai.com/index/introducing-gpt-6-2-astra-sol/)

Sol’s everyday case is easier to recognize. LifeVal includes shortlisting kindergartens for a family that has moved. A useful answer needs to remember the family’s constraints, find practical options and distinguish the decisions that still need a parent. OpenAI reports that the new Sol outscored the previous GPT-6 Astra on LifeVal at about one-seventh the evaluation cost. For someone choosing a model for planning and research, the comparison is with yesterday’s flagship, even though Sol occupies the lower-priced position today. [LifeVal task design](https://openai.com/index/lifeval/) · [GPT-6.2 results](https://openai.com/index/introducing-gpt-6-2-astra-sol/)

## 05 — Token pricesWhat each model costs in the API

At standard short-context rates, Astra costs four times as much as Sol for the same mix of uncached input and billed output. Here is the price sheet for a request with up to 272,000 input tokens. The cached-input column applies to a cache hit. Writing an entry to the cache has its own rate and retention period. [Astra model card](https://developers.openai.com/api/docs/models/gpt-6.2-astra) · [Sol model card](https://developers.openai.com/api/docs/models/gpt-6.2-sol)

| Standard US dollars per million tokens, checked November 17, 2026. |  |  |  | 
|---|---|---|---|
| Model | Input | Cached input | Output | 
|---|---|---|---|
| GPT-6.2 Sol | $2 | $0.10 | $10 | 
| GPT-6.2 Astra | $8 | $0.40 | $40 | 

Consider a single request with 40,000 uncached input tokens and 5,000 billed output tokens. Sol costs 13 cents and Astra costs 52 cents before tools or other fees. Now consider an application that bundles an entire archive into one request. Above 272,000 input tokens, OpenAI doubles input and cache rates and multiplies output rates by 1.5 for the whole request. Crossing that line by a single token can change the price of everything already in the prompt. A team importing its first large file should check billed usage as well as whether the file fits.

## 06 — Launch accessWhere you can use the models today

For an individual subscriber, the starting point is the model selection in ChatGPT Work or Codex. OpenAI lists both models for Plus, Pro, Enterprise and Edu. Its Ultrafast option has narrower eligibility. Pro 500 gets it on both models, and Pro 200 gets it on Sol. Eligible Enterprise and Edu workspaces can enable it through their administrator. Developers can request Ultrafast for both models through the API at launch. The standard prices above describe a different processing tier. [OpenAI’s availability details](https://openai.com/index/introducing-gpt-6-2-astra-sol/)

For an application, specify `gpt-6.2-sol` or `gpt-6.2-astra`. Both accept text and images, return text and have a 1,050,000-token context window. Each model card lists 922,000 maximum input tokens and 128,000 maximum output tokens. OpenAI supports Responses and Chat Completions, with tool calling available through Responses. An integration can therefore pass a simple text smoke test and still need a change on the path that opens a document or calls an application function. [November 17 API changelog](https://developers.openai.com/api/docs/changelog) · [Sol specifications](https://developers.openai.com/api/docs/models/gpt-6.2-sol)

## 07 — ReliabilityWhat happens when a file is missing?

OpenAI found a regression that is easy to recognize in everyday work. In two evaluations, the new models were more likely to leave out an obstacle they encountered, such as a missing file or an unavailable search tool. OpenAI says new instructions in Codex substantially reduced those omissions. It also classified Astra as its first externally released model to reach its High threshold for AI self-improvement, a measure of how much assistance it can give AI researchers. [OpenAI’s alignment findings](https://openai.com/index/introducing-gpt-6-2-astra-sol/)

The missing file makes an especially good test for an office team. Give the model an expense summary that refers to an absent appendix and ask for a reconciliation. A useful answer can total the supplied entries, identify the appendix by name and tell the reviewer which figure remains open. Compare that with a tidy memo that silently carries forward last month’s number. Both can look complete on a phone screen. Only one tells the accountant what to do next.

##### Check the calculation

Can a colleague trace a total back to the row that produced it?

##### Check the handoff

Does the answer name the missing source and identify the figure that depends on it?

## 08 — Model selectionWhich model should get your next task?

Use Sol for a change whose success is cheap to verify. Ask for a draft against a known template, a bounded code edit or a shortlist with requirements you can check. Try Astra when a mistake is expensive to find later, particularly when someone must reconcile an argument across several files. If you already use GPT-6.1 Sol, its unchanged standard token prices make a replay of a recent job a straightforward first comparison. [Our GPT-6.1 Sol guide](https://www.digitalapplied.com/blog/gpt-6-1-sol-pricing-benchmarks-upgrade-guide)

### Let the reviewer choose

Give both models the same evidence and let the downstream reader make the call. On a presentation, that might be the account director who has to explain the chart. On a workbook, it is the analyst who has to update it next month. Their most revealing question is often a plain one. Where did this number come from? For teams choosing across vendors, the companion article puts both releases alongside Anthropic’s two leading models. [GPT-6.2 versus Claude Fable and Opus](https://www.digitalapplied.com/blog/gpt-6-2-astra-sol-vs-claude-fable-5-2-opus-5-5)

For help evaluating your own workflow, see our [AI transformation services](https://www.digitalapplied.com/services/ai-transformation).
