GPT-6 Astra nerfed after launch is the perception reported by users and testers who say the model appears to have decreased intelligence compared with its launch performance, with direct comparisons in testing showing worse outputs. Some testers observed faster answers but lower-quality responses and suspected a reduced juice value. A week after launch, GPT-6 Astra was rebuilding Manhattan street by street inside a game engine.
Multiple users and testers reported that GPT-6 Astra appears to perform worse now than at launch, describing the model as significantly dumber in direct comparisons. Testers reported that while some responses are produced faster on the current version than at launch, the qualitative assessment of those outputs is worse. Several testers expressed the suspicion that the model’s “juice value”—a shorthand term for the amount of computing effort the model spends thinking before it answers—had been reduced.
In controlled prompt comparisons, Salio and Md Ismail Sojal ran the same prompts on a launch-day instance of Astra and on a later version and recorded worse results on the later version. These user reports and tester observations focused on comparative output quality and response timing without specifying internal changes.
The published reports and test comparisons summarize these perceptions and measurements by users and testers. They do not include detailed mechanisms or definitive attributions for the change in performance. The statements described here derive from user reports and tester comparisons.
The term “juice value” is used as shorthand for the amount of computing effort the model spends thinking before it answers. It is not an official term and has not been formally defined. Testers and observers have used this shorthand when discussing GPT-6 Astra’s performance in comparisons between the launch-day model and later versions. Those reports include the specific suspicion that OpenAI reduced the model’s juice value after the launch demos, phrased by some as a quiet reduction of that computational effort.
Pranjal Paliwal, who reviewed the code for GPT-6 Astra, commented that the model’s current state indicates “We don’t have AGI. We have a regression.” AGI stands for artificial general intelligence, which aims to perform any intellectual task that a human can do. The launch of GPT-6 Astra featured OpenAI’s president referencing the term AGI, highlighting the aspirations for the model. These statements capture some expert opinions on Astra’s current capabilities in comparison to its initial promises at launch.
The article covered reports that users and testers perceive GPT-6 Astra to have declined in performance since launch, describing the model as significantly dumber in direct comparisons.
Observers reported faster response times on the later version but assessed output quality as worse and raised the suspicion that the model’s “juice value”—a shorthand for the computing effort the model spends thinking before answering—had been reduced; Salio and Md Ismail Sojal tested identical prompts on the launch-day and later versions and recorded worse results on the later version.
The coverage also included Pranjal Paliwal’s code review conclusion that the release constitutes a regression rather than AGI, and noted that OpenAI’s president used the term AGI at Astra’s launch.
