AI roleplay
AI roleplay on the models you actually want
Pick the model, keep the story. Roleplay on Claude, GPT, DeepSeek, GLM and 35 uncensored models, and switch models mid-scene without losing the plot.
Chats stay on your device. No email needed.
Claude Opus 5- Claude Opus 5Anthropic
- Claude Opus 5.5Anthropic
Kimi K3Moonshot AI
Muse WonderMeta
GPT 6 AstraOpenAI
*I push open the tavern door, rain dripping from my cloak.* Any rooms left tonight?
Claude Opus 5The innkeeper eyes the storm, then your muddy boots. “One room, above the stables. A silver, and you tell me who’s chasing you.”
- AI models on NanoGPT
- 1,000
- Uncensored models
- 35
- Where chats are stored
- On your device
- Pro for heavy use
- $12/mo
What is AI roleplay?
AI roleplay is collaborative storytelling with an AI. You play one character and the AI plays everyone else: the narrator, the villain, the love interest, the world itself. You decide what your character does, the AI answers in character, and together you write a story neither of you planned alone.
Best models for roleplay
Ranked by NanoGPT’s live roleplay benchmark, which rewards emotional intelligence, creative writing and long-form consistency. Compare them on your own scene before you commit.
Best of these for emotional intelligence.
- Context
- 1M tokens
- 100 messages
- ≈ $2.50
Strong suit: emotional intelligence.
- Context
- 1M tokens
- 100 messages
- ≈ $2.00
Strong suit: vivid scene writing. Lowest price of these picks.
- Context
- 1.05M tokens
- 100 messages
- ≈ $0.33
#3 on NanoGPT for emotional intelligence.
- Context
- 1.05M tokens
- 100 messages
- ≈ $2.00
Also worth a try
Included in Pro, and #14 for roleplay on NanoGPT.
- Context
- 1.05M tokens
- 100 messages
- ≈ $0.22
From NanoGPT’s uncensored list, for scenes mainstream models refuse.
- Context
- 1M tokens
- 100 messages
- ≈ $1.10
Ranked by NanoGPT’s roleplay benchmark blend, with at most two models per maker. “100 messages” assumes about 3,000 tokens of story and character card sent with each message, and each model’s average reply length on NanoGPT (400 tokens where that isn’t known yet).
How it works
- Step 1
Pick a character
Choose one on the Characters page or write your own card.
- Step 2
Choose the model
Any model plays it, and you can switch mid-scene. Use an uncensored one for mature scenes.
- Step 3
Play the scene
Write what your character does. For long stories, turn on context memory.
Pick a genre and start
Copy a starter prompt into a character or any chat, or start the scene straight away.
Fantasy
“You are the narrator of a low-magic fantasy world. I’m a disgraced knight arriving at a border town at dusk. Describe what I see, then let me act.”
Sci-fi
“Run a hard sci-fi scene: I’m the night-shift engineer on a mining station when the reactor alarm goes off. Play the crew and the station AI.”
Mystery
“Be a 1920s detective story. I’m the inspector; you play every suspect at a country-house dinner where the host was poisoned. Keep the culprit secret.”
Romance
“Write a slow-burn romance scene. I play a bookshop owner; you play the regular who always buys poetry and never makes eye contact. Start on a rainy afternoon.”
Horror
“Start a quiet horror story. I’m house-sitting an old farmhouse and the phone keeps ringing at 3 a.m. Build dread slowly and let me choose what to do.”
Slice of life
“Play my new roommate on move-in day in a small apartment. Give them a distinct personality and a secret hobby I can discover over time.”
Text adventure
“You are the game master of a text adventure. Describe where I am in a few lines, list what I can see and do, then wait for my command. Track my inventory.”
Historical
“Set a scene in Venice in 1575 during the plague. I’m a young glassmaker’s apprentice; you play everyone I meet. Keep the history accurate and the stakes personal.”
“Start this scene” sends the prompt as your first message to your default model.
Keep long stories consistent
Long roleplays break in predictable ways. Here is what causes each problem and how to fix it on NanoGPT.
| Problem | Cause | Fix |
|---|---|---|
| The AI forgets earlier chapters | The story outgrew the model’s context window, so the oldest messages fall out. | Turn on context memory, which condenses older chapters, or switch to a longer-context model mid-story. |
| A character drifts out of voice | The character description is buried under hundreds of messages. | Keep a short character card or system prompt with their voice and limits; it is sent with every message. |
| Replies get repetitive | One model settles into the same phrases. | Switch models for a few turns, or regenerate. The story carries over because it lives in the chat, not in the model. |
| The model refuses a dark scene | Mainstream models filter violence and mature themes. | Use an uncensored model for that scene, then switch back. |
| Replies are too long or too short | The model guesses the length you want. | Say it once in the card or system prompt: “2 to 3 paragraphs, end on a choice for me”. |
| The story costs more as it grows | Every message resends the story so far. | Use a cheaper model for quiet scenes, or Pro for open models within its weekly allowance. |
Bring your own frontend
Want a dedicated roleplay app? Try SillyNano, NanoGPT’s own. Already roleplay in SillyTavern or Janitor AI? Use NanoGPT as the model provider and keep your characters, lorebooks and settings.
SillyNano
Open SillyNano →NanoGPT’s own roleplay app, with character cards, lorebooks, a scene tracker and Discover characters, on your NanoGPT models.
SillyTavern
SillyTavern setup guide →Connect SillyTavern to NanoGPT’s OpenAI-compatible API and pick any model from its list.
Janitor AI
Janitor AI setup guide →Use the proxy settings in Janitor AI to run your characters on NanoGPT models instead of the built-in one. Not sure which model? Read the best API for Janitor AI.
NanoGPT vs Character AI vs Janitor AI’s built-in model
| NanoGPT | Character.AI | Janitor AI (built-in JLLM) | |
|---|---|---|---|
| Model choice | 633 text models, 35 of them uncensored | Character.AI’s own models | One built-in model, or bring your own API |
| Filters | You pick mainstream or uncensored models | Strict; pornographic content is against its terms | Set by Janitor AI |
| Price | Pay per message, or $12/mo Pro | Free, or c.ai+ at $94.99/yr | Free |
| Where chats are stored | On your device | Character.AI’s servers | Janitor AI’s servers |
| No sign-up form or email needed | Yes | No | No |
Sources: c.ai+ pricing, Character.AI content rules and Janitor AI’s JLLM help page.
Frequently asked questions
Is AI roleplay on NanoGPT free?
You can try a free model without signing up. Better models cost per message, from a fraction of a cent, with top-ups from $0.10 in crypto. Heavy roleplayers can take Pro for $12 a month, which covers 60M input tokens a week on open models.
Which model is best for roleplay?
Right now Claude Opus 5 is the highest-ranked model on NanoGPT in its roleplay ranking, which blends EQ-Bench 4, EQ-Bench Creative Writing and EQ-Bench Longform Writing. The order changes as new models ship, so this page updates from the live data. See the full ranking.
Does the AI remember earlier chapters?
It remembers whatever fits in its context window, up to 1.05M tokens on today’s top roleplay models. For very long stories, turn on context memory, which condenses older chapters so the key facts stay in view, or switch to a longer-context model mid-story. How context memory works.
Are my roleplay chats private?
Your chat history is stored on your device, not on NanoGPT’s servers, and NanoGPT doesn’t log your prompts or replies. No email is needed. The model provider sees each message to answer it; turn on Zero Data Retention routing in Settings to use only providers that say they keep no copy. What is stored where.
Can I roleplay mature themes?
Yes, as an adult, with an uncensored model for scenes mainstream models refuse. Content that is illegal, including any sexual content involving minors, is never allowed and breaks NanoGPT’s terms. Uncensored models.
How much does a long session cost?
Each message resends the story so far, so cost grows with length. On Claude Opus 5, 100 messages cost about $2.50 with about 3,000 tokens of story, and about $16.00 once the story reaches 30,000 tokens. On GLM 5.3 the same long story costs about $1.84 per 100 messages, and Pro covers it within the weekly allowance. Context memory condenses old chapters so each message resends less; it is billed separately and not covered by Pro.
Is NanoGPT a good Character AI or AI Dungeon alternative?
Yes, if you want to choose the model. Character AI and AI Dungeon decide which models you get and how they are filtered. On NanoGPT you pick from 633 text models, including 35 uncensored ones, pay per message instead of a subscription, and keep your chats on your device. The text adventure starter above plays like AI Dungeon: the AI narrates, and you type what you do.
Your story, on the model you want
Pick a character or write your own, then switch models whenever the scene calls for it.