GPT-6 Astra’s One-Week Performance Collapse Sparks Nerfing Claims

OpenAI’s premier frontier model, GPT-6 Astra, drew widespread acclaim after early demonstrations showed it autonomously reconstructing entire sections of Manhattan street-by-street inside a game engine. The public debut was followed only one week later by frustration among software developers and enterprise users, who reported a sharp decline in output quality.

The complaints have focused on a familiar artificial intelligence industry problem: the cost of operating advanced “reasoning” models. Users across social media and developer forums accuse OpenAI of quietly “nerfing,” or downgrading, Astra’s cognitive capabilities to ease infrastructure strain and reduce operating costs.

Astra’s premium pricing has intensified that criticism. The model costs $10 per million input tokens and $50 per million output tokens, making it 2.5 times more expensive than the launch price of its predecessor, GPT-5.6 Sol.

Dax Raad, whose team develops the software development tool Opencode, said his company has returned to GPT-5.6 Sol. Raad reported that the team’s API expenditures doubled while using Astra, without producing benefits that justified the additional cost.

The model’s high-stakes positioning is sharply different from the everyday programming problems users describe. Astra is the first model in OpenAI’s portfolio to officially cross the company’s “critical threshold” for cybersecurity risk.

OpenAI’s safety guidelines state that this classification means Astra can autonomously discover and chain together previously unknown software vulnerabilities without human intervention. Because of those capabilities, access to the model’s raw offensive cyber tools is restricted to vetted defensive cybersecurity specialists through OpenAI’s proprietary Daybreak program.

Developer Pranjal Paliwal initially praised Astra’s coding ability, but said that a close examination of code generated a week after launch uncovered significant flaws. “We don’t have AGI,” Paliwal posted on X, referring to artificial general intelligence. “We have a regression.”

Other users have described a change in the model’s task processing. Developer Pankaj Kumar reported faster response times alongside a marked decrease in output quality, suggesting that OpenAI may have “reduced the juice value” of the system.

In large language model terminology, “juice value” is developer shorthand for inference-time compute: the processing power and time devoted to “thinking” through steps before a response is generated. For developers working at scale, those performance concerns have collided with the financial calculations of using Astra.

ChatGPT user Mustafa Sahinli compared Astra’s sudden decline in output to the backlash that followed Anthropic’s Claude Opus 4.6 one week after its release. Researchers Salio and Md Ismail Sojal separately ran identical prompts through launch-day Astra and the model’s current version, with both reporting inferior results from the newer sessions.

Another user, Saba, asked OpenAI why she now had to simplify and “dumb down” her prompts to make Astra execute tasks successfully. OpenAI has not issued a formal statement addressing the complaints.

The pattern is familiar to industry observers. In July, OpenAI faced an identical backlash after GPT-5.6 Sol rolled out and users said its advanced reasoning modes had suddenly become shallow. OpenAI executive Tibo Sottiaux denied at the time that the company had intentionally weakened the model, while confirming that engineers had been experimenting with “reasoning effort,” the dynamic setting that determines how many processing steps a model takes before responding.

AI companies frequently use optimization techniques to manage the immense financial and computational demands of operating frontier models at scale. One such method is quantization, which reduces the mathematical precision of a model’s internal weights, such as by converting 16-bit floating-point numbers into 8-bit or 4-bit integers.

Quantization can substantially reduce memory footprints, hosting costs, and latency. It can also produce subtle degradations in complex reasoning, mathematical computation, and coding accuracy.

Some users dispute the idea that Astra has undergone a secret technical downgrade. A pseudonymous developer using the handle Antikythera argued that the model performs exactly as it did on launch day, attributing the backlash to the natural end of the launch-week hype cycle.

“It is as dumb as it was on launch,” Antikythera wrote, adding that Astra has always tended to write like a “bullet-point-addict” and behave lazily when faced with highly complex tasks.

Theo, founder of T3Chat, offered a similar explanation. He said Astra has high output inconsistency compared with rival models such as Claude Fable, producing either brilliant or highly flawed code from the same prompt.

According to Theo, once the initial novelty faded, users simply began sharing Astra’s failures more frequently, creating the public perception of a sudden technical decline.