I've reached the peak of role-playing and completely lost interest. I want to share my experience: I got the best results on Chub, where there are more diverse bots. Janitor allows the use of free APIs (Mistral, Groq, Cerebras), but keep in mind: using your Groq key might get you banned for breaking the rules. The advantage of Chub over Janitor is the ability to create multitask prompts and customize character descriptions. On Chub, you can achieve a character staying in character 99.5% of the time, which you can't do on Janitor.
Right now, the most impressive language models are GPT-5 Chat and Claude Sonnet 3.7. ChatGPT-5 leads in character development and descriptions. Based on personal experience, it's easy to jailbreak without special methods, and the quality of its intimate scene descriptions is superb.
If you're interested in erotic descriptions, like those in romance novels, then Gemini 2.5 Pro and Gemini Flash are your good friends. System prompts that work great for ChatGPT-5 and Claude 3.7 don't work for Gemini. I often ran into Content Filters. Keep in mind: you could lose money and not get the story you want if you go beyond what's allowed in Google LLMs. However, their pricing policy is significantly more lenient than Anthropic's.
Models like mini can create erotic stories on the level of Gemini Flash, but they refuse to generate explicitly sexual content. Flirting and soft erotic undertones are permissible. At the same time, OpenAI Mini models, including ChatGPT 5 Mini, are among the most affordable in terms of price.
Of the free models, you should use DeepSeek v3 0324, Qwen3 A22 235b 0527, and Kimi K2 (content filtered as well but workaround is to reroll) from Moonshot AI. I've heard good things about GLM 4.5.
Most of the mentioned models are available for free on OpenRouter. If Kimi K2 refuses to generate content, delete the message and reroll. After 4-5 attempts, you will get the desired result.
I cannot give specific advice on sampling parameters. API vendors like Mistral, Groq, Cerebras often do not support all the necessary parameters. Rule of a thumb: when using paid ChatGPT and Claude models, you can increase the temperature to 1.20 without issues, while maintaining a high frequency penalty close to 1 and up to 1.10. It's not straightforward with open models. DeepSeek V3 and Kimi K2 usually work well at a temperature of up to 1.10. Qwen is unstable at that temperature; it's better to use it at 0.7-0.9.
