The use of LLMs for roleplay has long been treated as a fringe use case. After all, LLMs are promoted as tools to help with real work, like programming, automation, research, and more. There’s also a negative public perception of using LLMs for roleplay. It is often lumped together with AI companionship due to overlapping themes, though the hobbies are distinct.
OpenRouter’s State of AI paper analyzed over 100 trillion tokens of real-world LLM interactions across tasks, geographies, and time. The paper made one thing clear: AI roleplay is no longer a niche, but there’s a catch.
AI Roleplay Is One Of The Largest AI Use Cases
The paper analyzes metadata collected from the OpenRouter platform and reflects real-world activity on it. It states that “OpenRouter’s global scale and diversity make it a representative lens on large-scale LLM usage dynamics.”
According to the paper, “the sheer volume of roleplay and entertainment-oriented usage” was a “surprising finding.” AI roleplay accounted for roughly 52% of all OSS (open-weight) model token usage.
When looking at token usage for both OSS and proprietary models, AI roleplay is the second-most dominant category, surpassed only by programming.
DeepSeek’s models are the most popular OSS choice for AI roleplay, while Google’s models dominate among proprietary options.
AI Roleplay Takeaways
The paper presents a lot of data, and you can read it here. Below are the main takeaways related to AI roleplay.
- Nearly 60% of roleplay tokens fall under their “Games/Roleplaying Games” category, suggesting that users treat LLMs less as casual chatbots and more as structured roleplaying or character engines.
- This is further reinforced by the presence of Writers Resources (15.6%) and Adult content (15.4%), pointing to a blend of interactive fiction, scenario generation, and personal fantasy.
- Contrary to assumptions that roleplay is mostly informal, the data show a well-defined and replicable genre-based use case.
- Roleplay’s usage volume is immense, nearly rivaling programming. This is a striking insight: a consumer-facing roleplay application drives a volume of engagement on par with a top-tier professional one.
- AI roleplay falls under the low-cost, high-volume usage zone.
Why OSS Models Dominate
OSS models better cater to the AI roleplay community’s specific requirements. Roleplay is a hobby and a source of entertainment, similar to gaming. But not everyone is ready to pay for it yet, especially in regions where the currency is weaker compared to the US dollar.
Cost Factor
OSS models are cheaper to use compared to proprietary models, especially in longer conversations with large context windows. AI roleplay conversations can span hundreds of back-and-forth messages. Cost is a big factor behind people choosing OSS models that offer higher context limits at low input and output token costs.
- AI roleplay isn’t mission-critical. It doesn’t require the accuracy or reliability that expensive proprietary models guarantee.
- Unlike programming or other productive tasks, AI roleplay doesn’t justify premium pricing. There’s no investor funding or projected returns with AI roleplay.
- Users are willing to work around an OSS model’s flaws and come up with creative solutions to tackle them.
Censorship
Proprietary models often refuse to engage with mature or serious topics like war and violence, gore, horror, psychological thrillers and dramas, politics, and more. AI roleplay requires models that do not break character due to aggressive censorship or filters. Censorship and refusals hurt even more when you are literally paying the LLM to reject your request or give you a moral lesson.
- Those who use proprietary models for AI roleplay often use jailbreaks to circumvent the model’s safety alignment. However, using jailbreaks violates several providers’ Terms of Service. Companies like OpenAI and Anthropic frequently ban users who use jailbreaks to generate content that goes against their ToS.
- Thanks to platforms like OpenRouter acting as middlemen, users can use proprietary models for AI roleplay without the risk of having their accounts banned, unlike when using first-party APIs.
- Companies constantly combat jailbreaks, and circumventing filters is a never-ending game of cat and mouse.
Compared to proprietary models, most OSS models aren’t “safetymaxxed.” It’s easy to work around refusals with prompts/instructions. OSS models can also be fine-tuned for AI roleplay and to reduce or completely remove censorship.
Reliability
AI roleplay is a diverse hobby, with everyone having personal preferences and different expectations from LLMs. While one user may find a particular model’s performance “peak,” others may find it underwhelming. Models also behave differently to users’ prompts, sampler settings, and input.
Once someone finds a model they are happy with, they tend to stick with it. They may experiment with other models, but their “daily driver” remains the same. OpenRouter’s paper presented this retention pattern as the Cinderella “Glass Slipper” effect.
Compared to proprietary models whose only providers are their first-party APIs, anyone can provide inference for OSS models (as long as their license permits it). If you have the required hardware, you can run them locally too. The risk of “losing” your daily driver is less with OSS models.
But Here’s The Catch
While OpenRouter’s scale and diversity make the paper’s findings compelling, there are a few possible reasons that could have driven a spike in roleplay usage on its platform.
- OpenRouter, through its providers, offered limited free inference to several OSS models. The platform continues to provide free inference for select models.
- Regular users get 50 free daily requests, while those who deposit $10 into their OpenRouter account get 1000 free daily requests to the designated free models.
Free inference drove a lot of users from online AI roleplay platforms like JanitorAI and Chub to use models through OpenRouter on their preferred frontends.
- Community-created guides helped users easily “bring their own models,” and free inference allowed users try larger LLMs to enhance their experience.
- Several users even created multiple accounts to bypass the 50 free daily requests limit, rather than pay OpenRouter $10 for 1000 free requests a day.
Free inference continues to be the driving force behind current models with high roleplay usage. For example, free tokens account for more than 95% of TNG: DeepSeek R1T2 Chimera’s monthly usage.
OpenRouter no longer has any provider offering free inference for models like DeepSeek V3 and R1. While they were available, free tokens accounted for the majority of their usage, and they were primarily used for AI roleplay.
May Not Reflect Wider, Real-World LLM Usage
Free inference, word of mouth, community-created guides, and easy integration with AI roleplay platforms drove more users interested in AI roleplay to OpenRouter compared to other services. Inference providers were even forced to rate-limit free usage through OpenRouter to tackle service degradation and overload.
OpenRouter’s data may not accurately reflect wider, real-world LLM usage. Even if free tokens were excluded from analysis, the role free inference played in attracting users interested in AI roleplay cannot be ignored. While AI roleplay is no longer a niche, the use case isn’t as dominant as OpenRouter’s paper makes it out to be.
Still, AI Roleplay Is No Longer A Niche
AI roleplay is a new form of entertainment; it’s interactive and a completely customizable experience. So “the sheer volume of roleplay and entertainment-oriented usage” doesn’t surprise us or anyone interested in this hobby. However, to researchers and model developers, it is a “surprising find,” as noted in OpenRouter’s paper.
New online AI roleplay platforms pop up every day, many with the sole intention of making a quick buck and taking advantage of the growing interest. Established and reliable platforms are growing faster than expected. Even inference providers like Chutes are stepping into the space, with their AI roleplay platform, Fictio, now in closed alpha. Genuine projects with dedicated developers gain traction and fame and have an unbelievable amount of usage.
AI roleplay is not just casual chatbot interaction, as it is often perceived. Usage data shows that users create long-form narratives with defined scenarios, characters, and continuity. OSS models are currently the go-to for those interested in the hobby due to their lower cost, lack of censorship, and reliable availability.
While OpenRouter’s data may not accurately reflect wider, real-world LLM usage, it still shows that AI roleplay is no longer a niche. There’s a lot of demand for this use case, and it’s only growing.
