OpenAI Grayscale-Tests GPT-5.6-sol: Compute 'Halved'
According to MarsBit, OpenAI has been grayscale-testing a new version called GPT-5.6-sol on the Codex platform. The key change is a drastic reduction in inference compute resources—the Juice value dropped from 768 to 128, an 83% decrease. Juice is OpenAI's internal metric for measuring model inference computation; this adjustment means the model consumes significantly fewer resources when generating responses, thereby lowering operational costs.
Anthropic Accused of Silently Downgrading Claude Opus 4.8
Anthropic is also facing a user trust crisis. A large number of users have reported that Claude Opus 4.8 has recently shown obvious degradation in core capabilities such as logical reasoning and context memory, suspecting that the company has performed a silent downgrade (i.e., an unannounced version rollback). Unlike OpenAI's explicit compute cut, Anthropic's downgrade has not been officially confirmed, but test data from the user community has sparked widespread discussion.
Motivations: IPO Hurdles and Capital Winter
Industry observers generally believe that both AI giants adopting covert performance reduction strategies share common financial pressures. OpenAI and Anthropic have recently faced stalled IPO processes, and the capital winter in the AI sector has made fundraising more difficult. Meanwhile, the high cost of compute for large-scale training and inference models is forcing companies to seek cost-cutting measures. However, reducing service quality without user consent is severely damaging user trust.