OpenAI Grayscale Tests GPT-5.6-sol: Juice Value Plummets from 768 to 128
User reports indicate that OpenAI is conducting a grayscale test of its new GPT-5.6-sol model on the Codex platform. The most notable change is the drastic reduction of the inference compute metric (Juice value) from 768 to 128—a drop of over 83%. Juice value directly impacts the model's capability in complex reasoning, multi-step logic, and code generation tasks. This cut means OpenAI is sacrificing inference quality to reduce per-call costs, but user benchmarks show a significant rise in error rates on math, programming, and other hard tasks.
Claude Opus 4.8 Accused of Silent Downgrade: Logic and Memory Deterioration
Concurrently, Anthropic's flagship model Claude Opus 4.8 has come under fire. A large number of users on social media and forums have observed system-wide decline in logical reasoning, long-context recall, and response consistency—even forgetting details from earlier turns in conversations. The community suspects Anthropic of performing a silent downgrade without public disclosure, likely to reduce inference costs. Like GPT-5.6-sol, Claude Opus 4.8 may have had its allocated compute actively curtailed to cope with the soaring cost of inference.
Underlying Concern: IPO Blockage, Funding Winter, and Compute Cost Pressure
The near-simultaneous revelation of performance cuts by both AI giants is no coincidence. Industry insiders report that both OpenAI and Anthropic are facing impediments to their IPO timelines—OpenAI due to valuation disputes and Anthropic due to the sluggish fundraising environment. Compounding the problem is the global funding winter, which makes raising capital more difficult, along with persistently high rental prices for high-performance GPUs (e.g., H100, B200). As a result, both companies have resorted to lowering inference compute per query to control operational costs. However, such covert reductions damage user experience and erode market confidence in the reliability of AI services. If no proper explanation or performance recovery materializes, users may accelerate migration to alternative models or self-hosted inference solutions.