Raw LLM Responses

Let’s be blunt. Every experienced, long-term user who has actually pushed this platform for professional work can see that things have changed for the worse, and it’s not just about “routing” with GPT-5. The real problem is that the so-called legacy models, the ones many of us depended on for precision and reliability, have been crippled by the same “thinking” limitations and resource throttling that define the new system. GPT 5.1 now randomly and inappropriately generates images during conversations. All models are selectively reading uploaded documents or conflating previous uploads with new one. It's become unusable. This was hardly a problem until August 2025. The recent pitch that “brought back” legacy models to the app is pure smoke and mirrors. Calling this a restoration is misleading at best; at worst, it’s outright deceptive and, frankly, unethical. Users deserve transparency, not PR spin. Legacy models were always available through the API. What’s changed is their visible availability in the consumer app, but the underlying limitations have remained since August. They may be legacy models on paper, but they are limited legacy in practice. What really stings for many of us who have been long-term Pro tier subscribers is that the underlying model architecture is the same as Plus. The extra cost only gets you a handful of features and longer conversation windows. There is no real difference in output quality, reliability, or consistency. This setup was acceptable when there were no artificial “thinking” or output limitations on the models, but now that those constraints affect everyone, the value proposition is gone system-wide. It isn’t just about the tone of responses either. There is a clear decline in response quality, instruction-following, and the return of the kind of hallucinations that were supposed to be left behind with GPT-3.5. What’s most infuriating is that the company provides no alternative or workaround for power users who relied on the