GPT-6 Astra Users Say OpenAI's Newest Model Got Dumber. It Happened Before, Too

A wave of user reports and side-by-side tests have prompted accusations that GPT-6 Astra has lost capability shortly after its public launch. Early demonstrations showed the model performing complex creative and coding feats, but within a week many users on X began posting examples they said reveal sharper drop-offs in quality: slower or more superficial reasoning, worse code, and answers that feel less reliable than at launch. Several developers who initially praised Astra later posted direct comparisons or inspected the model's output and concluded performance has regressed. Complaints coalesce around a few recurring themes: responses that appear faster but lower in quality, a sense that the model is doing less internal computation — colloquially framed as a lower "juice value" — and specific reproducible prompt tests that returned worse outputs on the current Astra versus launch-day runs. Independent users including Md Ismail Sojal and others reported running identical prompts and finding degraded results. Some teams have reverted to using GPT-5.6 Sol because they judged Astra's higher cost did not justify the trade-offs in output consistency. Not all observers attribute the shift to a discrete throttling or deliberate weakening by OpenAI. Some argue the initial enthusiasm masked shortcomings that became more visible once more people performed detailed tests; others say Astra's behavior is simply more inconsistent than alternatives, producing both exceptional and poor outputs depending on the prompt. OpenAI previously faced a similar cycle in July with Sol, when executives confirmed the company was experimenting with the model's "reasoning effort" setting while denying intentional weakening. Technical explanations floated by users include changes in model precision (quantization) or adjustments to internal reasoning effort, though OpenAI has not publicly confirmed deliberate quantization or a policy of downgrading a deployed model. Astra remains notable for being OpenAI's first model identified as crossing a "critical threshold" for cybersecurity risk, and access to its most powerful capabilities is restricted under the company's Daybreak program. The model's pricing also stands higher than Sol's launch price: Astra charges $10 per million input tokens and $50 per million output tokens, about 2.5 times Sol's initial rates, a factor some users cite when deciding whether to continue using it.