OpenAI The new model GPT-6 Astra has encountered a lot of user skepticism on social media platforms just one week after its release. Some developers claim that the model performed impressively at its launch, but in recent days, the quality of its output has significantly declined, especially in code generation and handling complex tasks, with a noticeable gap compared to the early stages of its release.

Users start posting comparison results.

Some users posted comparison screenshots, claiming that under the same prompts and settings, the results provided by the current version of Astra are not as good as on the day it was launched. Pranjal Paliwal, who had previously publicly praised the developer of that model, stated that after rechecking the code written by Astra, he found that the quality was not up to expectations, even describing it as a "regression" rather than progress.

Researcher Md Ismail Sojal also stated that he used the same prompt word in both the initial version and the current version, and the results showed that the latter performed worse. Other development teams have mentioned that they have reverted to the previous generation of models GPT-5.6 Sol, due to increased usage costs, but the benefits are not significant.

The controversy focuses on reasoning computing power.

Around this controversy, the topic that users discussed the most was the so-called “juice value”, which refers to how much computing resources the model uses for inference before answering. Although this is not an official term in OpenAI, many users suspect that the company may have reduced the model's inference intensity after the demonstration ended in order to control power consumption and operating costs.

There are also more direct speculations that the model may have undergone quantization processing. Quantization typically refers to reducing the internal computational precision of a model in order to cut costs and minimize resource usage, but this often comes at the expense of some accuracy. The report mentions that OpenAI has never confirmed that it would actively weaken the released version in this manner after the model goes live.

Some also argue that the problem is not a new one.

Not everyone agrees with the claim that "the model has been weakened." A user using the pseudonym Antikythera believes that there has been no essential change in the performance of Astra since its launch. The problem lies in the overheated market sentiment during the initial phase; users only began to test it more thoroughly after the hype subsided, which revealed inherent shortcomings of the model, such as lazy responses and a preference for bullet-point writing styles.

T3Chat founder Theo also expressed a similar view. He believes that the issue with Astra is more of a lack of stability: sometimes it can produce very outstanding results, while at other times it may make obvious mistakes. Therefore, the large number of negative cases that have appeared on social media recently could also be due to users sharing their failed outcomes more frequently after the "honeymoon period" has ended.

Sol has also experienced similar controversies.

This kind of controversy is not a first. Reports mention that the previous generation of flagship models GPT-5.6 and Sol on OpenAI also experienced similar issues in July this year. At that time, users claimed that their advanced reasoning modes became less sophisticated in a short period. OpenAI executive Tibo Sottiaux denied that the company intentionally weakened the models, but confirmed that the team did test a "reasoning effort" setting, which controls how many steps the model would consider before answering.

As of now, OpenAI has not made any public response to this round of controversy regarding Astra. The report also mentions that Astra is still recognized by OpenAI as the first model to overcome the critical threshold of network security risks, possessing the ability to discover and connect unknown software vulnerabilities without any assistance. Therefore, it is only available to approved defense parties.

Additional information:GPT-6 Astra The current pricing is $token per million inputs and $token per million outputs, which is approximately 2.5 times the price at the time of GPT-5.6 Sol.