Event Overview: Compute Cuts and Silent Downgrades
According to reports from MarsBit, OpenAI has been covertly testing a new version of its flagship model, tentatively named GPT-5.6-sol, on the Codex platform. This variant drastically reduces the inference compute budget — referred to as the 'Juice value' — from 768 on the standard GPT-5.6 to 128, a reduction of over 83%. Effectively, this slashes the model’s capacity for complex logical reasoning, long-context memory, and multi-step problem solving. Simultaneously, Anthropic’s Claude Opus 4.8 has come under fire from users who claim to have detected a silent downgrade. Without any official announcement, the model’s performance in tasks such as mathematics, coding, and context retention has reportedly deteriorated significantly. Users compared recent outputs with earlier sessions and found higher error rates, broken contextual chains, and even forgotten instructions mid-conversation.
Behind the Scenes: IPO Hurdles and Capital Winter Driving Cost Reduction
Market observers argue that the timing of the two companies’ hidden performance cuts is no coincidence. OpenAI is currently navigating an IPO process, with investors increasingly concerned about its ballooning operational costs, particularly compute expenses. Anthropic, meanwhile, faces its own fundraising challenges amid a broader capital winter, with recent funding rounds falling short of expectations. With the cost of training and inference for large language models remaining exorbitant, reducing inference-stage compute is the most direct lever for cost control. The GPT-5.6-sol version, for instance, could lower per-inference costs by over 80%. However, this cost-saving measure implemented without user notification severely undermines trust—especially among paying subscribers of ChatGPT Plus and Claude Pro, who now receive degraded performance for their money.
Industry Implications: Transparency and Ethical Concerns in AI Services
This incident sheds light on a growing issue in the AI industry: when models face commercialization pressures, companies may quietly adjust model performance without public disclosure—a practice akin to 'feature reduction without notice' in traditional software. The crypto and tech communities have responded strongly, with many developers reconsidering their reliance on OpenAI and Anthropic APIs in favor of open-source alternatives like Meta’s Llama or Mistral. On social media, users have started a 'refund' movement, demanding that both companies reveal version change logs and offer performance guarantees. Over the long term, if major model providers continue such opaque operations, the entire trust framework for AI services could be shaken, potentially prompting regulators to step in and mandate algorithmic transparency.