AI Card Bao AICARDBAO.COM

Industry News

OpenAI releases GPT-6.2 Astra and Sol: Astra’s price is reduced by 20%, Sol’s price remains unchanged, and Ultrafast is lowered to Pro 200 for the first time

Two weeks ago, it was just an extra line of names in the Codex usage details, but now it has become a regular one. OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol, one large and one small are online at the same time.

The division of labor is the same as the previous generation. Astra is the flagship, and OpenAI says it is prepared for the "most difficult projects"; Sol focuses on daily coding and writing, is cheap and runs fast.

The two models have been included in ChatGPT Work and Codex, and can be used by Plus, Pro, Enterprise, and Edu users. Developers can just adjust gpt-6.2-astra and gpt-6.2-sol in the API.

Open the model selector of Codex, GPT-6.2 Sol has reached the first place in the default recommendation, Astra ranks second, and the previous generation 6.1 Sol and 6 Astra have moved down. OpenAI itself said that many employees in the company use 6.2 Sol by default on a daily basis.

The API price is calculated as follows (per million tokens):

GPT-6.2 Astra costs $8 for input, $0.40 for cached input, and $40 for output, making standard input and output prices 20% lower than GPT-6 Astra. GPT-6.2 Sol has no price adjustment and continues to use GPT-6.1 Sol’s $2 input, $0.1 cache, and $10 output. Comparing the two, the Astra is exactly 4 times as powerful as the Sol in all three categories.

The unit price has only dropped by 20%, but the cost of a single task given by OpenAI has dropped much further: while setting new results in professional work and scientific research evaluations, the cost of each task is less than one-third of GPT-6 Astra. The difference is due to the shorter output. Officials say that the two new models use fewer output tokens to complete difficult tasks, and ChatGPT users can hit the usage limit later.

Sol's numbers are more interesting. On Cognition's FrontierCode 2.0, both open at the highest inference intensity, Sol and Astra are only half a point apart, and the cost is about one-fifth of Astra. This test looks at whether changes to the real code base can be merged by maintainers, which is closer to daily work than answering questions.

These are the results announced by OpenAI itself, and third-party evaluations have not yet been released.

On the Astra side, OpenAI gave examples that focus on office and engineering: use a screenshot of a slide as a template and let it create a new page according to the layout, color matching and font size; early testers reported that the code style written in Codex is cleaner, the testing is more focused, and the number of rounds of back-and-forth changes is reduced.

Another change lands on Ultrafast. The speed of issuing tokens in this mode can be up to 4.5 times the standard speed, at the cost of more credits. Both the $500 Pro model can be driven; the $200 Pro model gets Ultrafast for the first time, starting with the 6.2 Sol. Enterprise and Edu need to be opened in the background by the administrator. It can also be used in the API.

The timeline is as follows: GPT-6 Astra was released in early September, and GPT-6.1 Sol took over on DevDay at the end of September. This time Astra skipped 6.1 directly to 6.2, and the version numbers of the two lines were finally aligned.

In the past, gpt-6.2-astra was revealed in the Codex bill in advance, and about 20% of the requests were grayscaled. We wrote an article "GPT-6.2 Astra has not been officially announced yet, Codex bill details were revealed first." Those who felt that the Codex results were good and bad during that time can now be corrected.

How to choose depends on the job. For daily code writing and document modification, 6.2 Sol is enough, and the price is a quarter of Astra; for big projects that are really stuck, switch to Astra. According to OpenAI's own data, Astra's single-task cost has dropped to less than one-third of the previous generation. Teams that originally thought GPT-6 Astra was too expensive can recalculate the calculation.

Related