GPT-6 Astra users say OpenAI's latest model has become stupid, a similar situation has occurred bef

GPT-6 Astra word-of-mouth reversal: Is it model degradation or sobriety after the "honeymoon period" ends?

A week ago, GPT-6 Astra was recreating Manhattan street scenes street by street in the game engine, and its superior ability amazed users. However, this week, the same group of users posted screenshots asking what OpenAI had done to the model.

Well-known anonymous developer synthwavedd posted on the X platform: "Today, Astra feels obviously stupid." He went on to sigh: "It's a pity that 'post-release lobectomy' will come sooner or later."

Familiar narrative: From stunning to "weakening"

This is a familiar phenomenon. AI users are used to their favorite chat robots or agents suddenly feeling "nerved." Typically, there is a reasonable explanation for this perceived performance degradation. But when enough people make the same complaints at the same time, it is necessary to take a look at what happened. Currently, social media is full of examples of users quickly losing their liking for GPT-6 Astra.

Pranjal Paliwal, the developer who had previously praised the model, went back and carefully examined the code Astra had written for him, and he was disappointed. "We have not implemented General Artificial Intelligence (AGI), we have just experienced a technological setback," he said in the post.

AGI (Artificial General Intelligence) is the term used by the industry to describe machines that are capable of cognitive operations like humans. The president of OpenAI used the term at the Astra conference. But a week later, a user used it to describe the exact opposite situation.

Common characteristics of complaints: "juice value" versus performance trade-off

These complaints take on a similar pattern. Developer Pankaj Kumar listed the symptoms: the answers were faster but of worse quality; and suspected that OpenAI had "reduced the juice value." Founder Saba directly questioned OpenAI why she now had to "de-smart" models to complete the task.

"Juice value" is not an official term and no one formally defines it, but here it is the tacit shorthand: it refers to the amount of thought and calculation the model puts into before answering. People suspect OpenAI quietly lowered this parameter after releasing a demo to complete its marketing mission.

Some people conducted rigorous tests. Salio and researcher Md Ismail Sojal both used the same prompt for the launch-day (release day) version of Astra and the current version, and both found performance degradation in the current version.

Others choose to return to the old version. Dax Raad, who built the coding tool Opencode, said that his team had returned to using Astra's predecessor model, GPT-5.6 Sol, because the shortcomings of doubling the cost were not worth it. ChatGPT user Mustafa Sahinli bluntly said that the current Astra makes him feel like Claude Opus 4.6 was "a week after its release"-implying that Anthropic's own products have experienced similar post-release boycotts.

Opposition: Not getting stupid, but fading away

However, not everyone believes OpenAI has deliberately weakened its latest model. Anonymous user Antikythera issued the most detailed rebuttal yet, arguing that the timeline had been reversed.

Antikythera wrote: "It's as stupid as it was when it was released. The model is good, but there are many problems. It's lazy and written like someone who relies on bullet points... people were overhyped during release week, and now they have time to test it and see its mistakes." [TAG

In other words: nothing has changed. You're just not dazzled so much anymore that you notice the model's ever-present glitch of listing bullets.

T3Chat founder Theo has a similar view: Astra is actually more unstable than Claude Fable, sometimes generating excellent code, sometimes extremely stupid. Now it seems that as the honeymoon period ends, users post those bad results more frequently.

Historical Reenactment and Conspiracy Theories

This situation is not new. OpenAI's last flagship product, GPT-5.6 Sol, went through exactly the same cycle in July, when users reported that their top-level reasoning patterns became shallow overnight. Tibo Sottiaux, an OpenAI executive, denied intentionally weakening the model, while confirming that the company had been trying to tweak "reasoning effort," a setting that controls the number of "thinking" steps the model takes before answering.

One reply summed up the current popular joke: "Closed Source Labs releases a model, and a few days later it seems to catch some disease and suddenly becomes stupid." Another netizen put forward the theory with a sarcastic remark: "They may have quantified the model to save costs so that it wouldn't burn so many companies 'money."

Quantifying a model means reducing the accuracy of its internal mathematical calculations to cut costs, but this usually comes at the expense of accuracy. OpenAI has never confirmed its intention to do this on published models.

Status and Costs

As of now, OpenAI has not issued a statement similar to the Sol period on Astra. The model remains the company's first to cross so-called critical thresholds of cybersecurity risks, which means it can discover and concatenate unknown software vulnerabilities without human guidance. Under OpenAI's Daybreak plan, such capabilities are limited to vetted defenders.

No matter how performance changes, its price remains high: $10 per million input tokens and $50 per million output tokens, which is 2.5 times the price when Sol was released.