It’s September 15, 2026, and OpenAI’s latest flagship model has a problem that anyone who’s ever lived through a software update knows all too well: the promise was big, but the reality falls short.
A week after the release of GPT-6 Astra, a growing chorus of complaints is getting louder. Developers who initially praised the model to the skies are now sharing screenshots and comparisons showing that output quality has declined. The term that keeps coming up: “post-launch lobotomy.”
“We don’t have AGI, we have a regression”
Developer Pranjal Paliwal, who praised the model at launch, went back and read the code Astra had written for him. His conclusion was devastating: “We don’t have AGI. We have a regression.”
Other users report similar complaints. Answers come faster, but quality is noticeably worse. Multiple developers, including Md Ismail Sojal, ran the exact same prompts on the launch-day version and the current version. Today’s results are demonstrably worse.
The popular theory among users is that OpenAI quietly lowered the “juice value” – an unofficial term for how much compute the model spends thinking before it answers. The suspicion is that the dial was turned back down once the launch demos had done their job.
Not everyone is convinced
There’s a counterargument too. An anonymous user named Antikythera published the most detailed rebuttal, with the core argument that the timeline runs backwards. “The model is just as dumb as it was on release day. The model is good, but it has a lot of problems. It’s lazy. Writes like a bullet-point addict.” According to this analysis, nothing has changed – people were simply too excited during launch week and are only now noticing the flaws.
T3Chat founder Theo has a similar view: Astra is just less consistent than Claude Fable 5.1, which means it sometimes produces brilliant code and sometimes incredibly dumb code. What’s happening now is that users are sharing bad results more often now that the honeymoon phase is over.
OpenAI stays silent
Notably, OpenAI has so far given no official response to the complaints. This despite the fact that in July, the company did respond when similar complaints surfaced about the previous flagship model GPT-5.6 Sol. At the time, an OpenAI director denied that anything had been deliberately throttled, but acknowledged that the company continuously experiments with settings for “reasoning effort.”
The timing of this controversy is painful. Earlier this month, OpenAI presented Astra as the model that would usher in the AGI era. President Greg Brockman used those exact words during the launch briefing. A week later, one of his own users is using the same acronym to describe the opposite.
Laat een reactie achter