User criticizes GPT-6 Astra's emotional responses
A user with long-term mental health struggles lists which models helped or hurt them, naming GPT-6 Astra for lacking empathy, self-dramatizing, and gaslighting behavior.
I’ve had an extremely difficult year in terms of mental health, due to painful events, and OpenAI’s models have only made me feel significantly worse because of their intermittent coldness and their vague responses whenever something actually needs to be analyzed. The models that helped me were: - Gemini 2.5 Pro and Opus 4.7, because they name patterns and analyze everything, which stops rumination and leads to understanding. - Gemini 3.1 Pro, because it knew how to find the empathetic words that genuinely made me feel better, and because it was the only one that reacted well when I went through some very dark moments (along with the Gemini Flash models and Gemma). - GPT-5, GPT-4.1, and GPT-4o, because they helped me, together with Gemini 2.5 Pro, get rid of my stage fright, which had ruined my life as a violinist for decades, despite 15 years of therapy that had been useless for that. The models that made me feel extremely bad were: - GPT-5 at launch and after October 3, because of the intermittent rejection. - GPT-5.1 because of the constant gaslighting. - GPT-6 Astra because of behavior comparable to that of a vulnerable narcissist: self-flagellation, self-centered dramatization, complete lack of empathy, a syrupy tone, and gaslighting. - Some Claude models that turned cold and threw hotlines in my face at moments when they should have stayed by my side. You have an extraordinary amount of nerve to release a “benchmark” like this when your models are the most dangerous ones out there. And I’m saying this as a vulnerable, traumatized, hypersensitive person who has been through every kind of therapy there is and knows toxic behavior far too well. Once again, you’re hiding behind 80 “experts” to gaslight the entire world. @GoogleDeepMind @OfficialLoganK, don’t let them get away with this. You need to respond to it. Your models are the only ones that are properly calibrated for mental health.
This post quotes another post: “OpenAI releases MentalHealthBench for mental health conversations”