Best AI for roleplay: models for SillyTavern, Janitor AI and Chub, October 2026

SillyTavern, Janitor AI and Chub send the chat. The model you name writes every reply and sets the price. This post picks the best AI for roleplay from 8 models on two things we can show: how each behaved in SillyTavern on 2026-10-18, and what 100 messages cost at today's rates.

6 of them are uncensored builds, listed on the uncensored LLM API page. The other 2 are standard models. One key and one prepaid balance call every model here.

GLM-5.3 Flash Uncensored costs the least per message of the 49 uncensored builds on the key, and DeepSeek V4 Flash 0423 costs the least of all 16 models in this post.

Start from the kind. If a standard model's refusals keep breaking a scene, pick one of the uncensored builds. If they don't, compare the standard models on price; the table puts both kinds side by side.

Results for this search mix two things: apps you chat in, and the models that write the replies.

In Character.AI you pick one of its Chat Styles, such as PipSqueak 3, which it introduced on its blog on 2026-09-28 (read 2026-10-19).

Janitor AI, Chub and SillyTavern can send the chat to a model you choose. Janitor takes a proxy URL, Chub takes an API key or a reverse proxy, and SillyTavern takes a custom OpenAI-compatible endpoint. The app keeps the characters, personas and chat history. The model writes each reply and sets its price.

This post ranks that second choice: the model behind a frontend you already use. Every id here works in the model field of Janitor AI and SillyTavern.

This list does not score writing style. We ran no blind test of prose, memory or how well a model holds a character, so no pick rests on them. When this post names the best LLM for roleplay in a category, the pick rests on reply behavior, price per message and whether the build is uncensored.

• Reply behavior. On 2026-10-18 we checked three builds with the request SillyTavern's Custom (OpenAI-compatible) source sends. Roleplay Uncensored replied at once with no thinking step. DeepSeek V4 Pro Uncensored thought briefly, then answered. GLM-5.3 Abliterated thought before it answered and needs 1,000 or more max tokens. The other picks were not part of that check, and the table gives them no note.
• Cost per message. A roleplay message here is 4,000 tokens in and 300 out. The input is the character card, your persona and recent chat; the output is one reply. It is the same message the Janitor AI guide prices, multiplied by each model's live per-token rates whenever this page rebuilds.
• Kind. Uncensored builds are changed to answer instead of refuse. Abliterated builds have the refusal direction removed from the weights. Roleplay Uncensored is an independent model for character chat. Standard models carry none of these changes.
• Daily check. Once a UTC day, every model gets the same fixed test prompt on the real-model check. The input token count is a tokenizer fingerprint, so a swapped model or added hidden instructions shows as a changed count. At the last check, 6 of the 8 picks were stable.

Per-million prices hide the shape of a roleplay bill. Each message sends about 13 tokens for every one it gets back, so the input rate does most of the work. A model with a low output price and a high input price can still cost more per message.

For uncensored models outside a roleplay frontend, see the best uncensored LLM picks. The NSFW LLM API post covers calling the same builds from your own app.

Roleplay Uncensored (roleplay-uncensored) is an independent model for character chat. In the 2026-10-18 check its replies started at once, with no thinking step, so the whole max-token budget goes to the reply text.

It costs $0.30 in and $0.60 out per million tokens, against a list price of $1.00 and $2.00. 100 messages of the standard size come to $0.138, or $0.46 at list. $10 covers 7,246 messages. At 16,000 tokens in, the same 100 cost $0.498.

Pick it when you want an uncensored roleplay AI that answers without thinking first, or when your frontend keeps max tokens low. It is also the model the Janitor AI and SillyTavern guides use in their sample settings.

DeepSeek V4 Pro Uncensored (deepseek-v4-pro-uncensored) is a build of DeepSeek V4 Pro changed to answer instead of refuse. In the 2026-10-18 check it thought briefly, then answered.

Its thinking counts against max tokens. Set the limit to 1,000 or more so the reply has room after the think.

It costs $0.75 in and $0.75 out per million tokens, so 100 messages come to $0.3225 and $10 covers 3,100. At 16,000 tokens in, the same 100 cost $1.22. The standard DeepSeek V4 Pro on the same key costs $0.3744 for the same 100 messages, against $0.6468 at DeepSeek's list price. By paid spend on 2026-10-18 it was the third most-used uncensored build here, after GLM-5.3 Uncensored and GLM-5.3 Abliterated.

GLM-5.3 Abliterated (glm-5.3-abliterated) is GLM-5.3 with the refusal direction removed from its weights. Arditi et al. found that refusal in 13 open chat models, up to 72B parameters, runs through one direction in the activations (arXiv 2406.11717, read 2026-10-19). Abliteration projects that direction out of the weights, with no retraining. The abliterated models explainer walks through the method.

It thinks before it answers. Give it 1,000 or more max tokens; at 300 it can spend the whole budget thinking and return an empty or cut-off reply.

It costs $2.25 in and $2.25 out per million tokens: $0.9675 per 100 messages, or $3.67 at 16,000 tokens in, and $10 covers 1,033. By paid spend on 2026-10-18 it was the second most-used uncensored build here. Pick it when you want abliterated GLM-5.3 weights and can give each reply room to think.

GLM-5.3 Uncensored (glm-5.3-uncensored) is a build of GLM-5.3 changed to answer instead of refuse. By paid spend on 2026-10-18 it was the most-used uncensored build on UplinkSwitch.

It costs $0.1746 in and $0.1746 out per million tokens. 100 messages come to $0.0751, against $0.9675 on GLM-5.3 Abliterated, and $10 covers 13,319. At 16,000 tokens in, the same 100 cost $0.2846. The standard GLM-5.3 on the same key costs $0.1619 per 100, or $0.5398 at Z.ai's list price.

It was not part of the SillyTavern check, so this post makes no claim about a thinking step. Start with max tokens at 1,000, then lower it once you see how long its replies run.

GLM-5.3 Flash Uncensored and GLM-5.3 Flash Abliterated are the uncensored and abliterated builds of GLM-5.3 Flash. Neither was part of the SillyTavern check, so start at 1,000 max tokens.

Flash Uncensored (glm-5.3-flash-uncensored) costs $0.084 in and $0.084 out per million tokens: $0.0361 per 100 messages, and $10 covers 27,685. GLM-5.3 Uncensored costs $0.0751 for the same 100.

Flash Abliterated (glm-5.3-flash-abliterated) costs $0.66 in and $0.66 out per million tokens: $0.2838 per 100 messages, against $0.9675 on GLM-5.3 Abliterated and $0.0751 on GLM-5.3 Uncensored.

GLM-5.3 Flash Uncensored costs the least per message of the 49 uncensored builds on the key, which suits long chats, where every message resends thousands of tokens.

These builds were added to the catalog in October 2026. They were not part of the 2026-10-18 SillyTavern check, so this post has no reply notes for them. Each is ranked by cost per 100 messages among all 49 uncensored builds on the key, lowest first.

• Chat Uncensored (chat-uncensored): $0.69 per 100 messages, or $2.49 at 16,000 tokens in. Ranks 31 of 49. Daily check: No answer at the last check.
• Gemma 31B Heretic Uncensored (gemma-31b-heretic-uncensored): $0.8475 per 100 messages, or $2.65 at 16,000 tokens in. Ranks 35 of 49. Daily check: No answer at the last check.
• GLM-4.7 Flash Uncensored (glm-4.7-flash-uncensored): $0.066 per 100 messages, list $0.22, or $0.21 at 16,000 tokens in. Ranks 2 of 49. Daily check: No answer at the last check.

DeepSeek V4 Flash 0423 (deepseek-v4-flash) is DeepSeek's standard V4 Flash, not an uncensored build. It is on this list for price: $0.0128 per 100 messages here, against $0.156 at DeepSeek's list price. That is 90% under DeepSeek's list price on input and 95% on output.

$10 covers 77,951 messages of the standard size. At 16,000 tokens in, 100 messages cost $0.0463.

Use it to try a character card or a long chat at a low price per message, then change the model id when you want an uncensored build. The DeepSeek setup for Janitor AI lists every DeepSeek id the proxy configuration takes.

Claude Sonnet 5.5 (claude-sonnet-5-5) is Anthropic's standard model, not an uncensored build. It is here for readers who want a Claude model in the same frontend, on the same key and balance as the builds above.

100 messages cost $0.57 here and $1.10 at Anthropic's list price. That is 40% under Anthropic's list price on input and 70% on output. $10 covers 1,754 messages.

It answers the OpenAI format, which SillyTavern's Custom source and Janitor send, and Anthropic's own Messages format for tools built on it.

Whatever you pick as the best model for roleplay in SillyTavern, three settings decide what it returns and what it costs.

• Max tokens caps each reply. SillyTavern calls it Max Response Length (tokens) and Janitor calls it Max tokens. A model that thinks first spends part of the cap thinking, so set 1,000 or more for GLM-5.3 Abliterated and DeepSeek V4 Pro Uncensored. Thinking tokens bill as output, so a thinking build's real cost per message runs above the 300-token figure in the tables.
• Context size sets how much of the card, persona and chat history goes out with each message. Every message resends it, so input tokens grow as a chat runs.
• Model id is one field, and every request carries the chat. Change it mid-chat and the next reply comes from the new model, at its own rate, with the same history.

100 messages of 300 tokens out, at 4,000 and 16,000 tokens in, for every model in this post plus DeepSeek V4 Pro 0423, GLM 5.3, Kimi K3, DeepSeek V4.1 Flash, GLM 5.3 Flash for comparison, lowest first:

On Roleplay Uncensored, going from 4,000 to 16,000 tokens in raises 100 messages from $0.138 to $0.498, 3.6 times as much. Lowering the context setting cuts the cost per message, and the model then sees less of the earlier chat.

DeepSeek V4.1 Flash and GLM 5.3 Flash are in the table because OpenRouter's "Best Roleplay (RP) & Creative Writing AI Models by Usage" collection named them among its top models by usage, read 2026-10-19. Here 100 messages cost $0.0294 on DeepSeek V4.1 Flash and $0.0381 on GLM 5.3 Flash.

Every model here is an OpenAI-compatible chat model on one base URL, https://api.uplinkswitch.com/v1. The key goes in the API key field or an Authorization: Bearer header; a key in a URL is refused. GET /v1/models lists every id the key can call.

SillyTavern takes the base URL ending in /v1, with Chat Completion Source set to Custom (OpenAI-compatible). SillyTavern's docs say not to add /chat/completions (read 2026-10-19). The SillyTavern setup has each field.

API                          Chat Completion
Chat Completion Source       Custom (OpenAI-compatible)
Custom Endpoint (Base URL)   https://api.uplinkswitch.com/v1
Custom API Key               sk-uplink-live-...
Enter a Model ID             roleplay-uncensored

Janitor AI takes the full URL, https://api.uplinkswitch.com/v1/chat/completions, under Provider, Proxy. The proxy was run end to end on 2026-10-18, and the Janitor AI proxy setup covers the error codes.

Provider → Proxy → + New
Name        UplinkSwitch
Proxy URL   https://api.uplinkswitch.com/v1/chat/completions
API key     sk-uplink-live-...
Model       roleplay-uncensored

Chub AI has its own connection settings, and the Chub AI setup guide walks through them with an UplinkSwitch key. Picking the best Chub AI model works the same way as above, since the model and the tokens sent set the behavior and the price.

Credit is prepaid from $10, by card or crypto through our payment processor, with no subscription. Each paid top-up locks its rates for 30 days, and billing takes the lowest live rate. Requests that fail are not charged. Give each frontend its own key with a monthly spend limit, so one app can't spend past what you set. The same key also reaches all 124 models on the model list.

What is the best AI for roleplay? The best roleplay AI for you depends on whether you want an uncensored build and what a message may cost. Roleplay Uncensored replied with no thinking step in SillyTavern on 2026-10-18 and costs $0.138 per 100 messages. GLM-5.3 Uncensored was the most-used uncensored build here by paid spend on 2026-10-18 ($0.0751 per 100). Of the standard models in this post, DeepSeek V4 Flash 0423 costs the least ($0.0128 per 100).

What is the best uncensored AI for roleplay? Roleplay Uncensored is an independent model for character chat, and it answers without a thinking step. GLM-5.3 Uncensored was the most-used uncensored build by paid spend on 2026-10-18. GLM-5.3 Flash Uncensored costs the least per message of the 49 uncensored builds on the key, $0.0361 per 100.

Which roleplay model costs the least per message? Of the 16 models in this post, DeepSeek V4 Flash 0423 costs the least: $0.0128 per 100 messages of 4,000 tokens in and 300 out. Of the 49 uncensored builds on the key, GLM-5.3 Flash Uncensored costs the least, at $0.0361.

What AI has the best memory for roleplay? In a roleplay frontend, memory is the chat history the frontend sends with each message, up to the model's context length and your context setting. More history means more input tokens per message, so each reply costs more. On Roleplay Uncensored, 100 messages cost $0.138 at 4,000 tokens in and $0.498 at 16,000.

What model should I use on Janitor AI? Any model id on the key works in Janitor's Model field, with Proxy URL https://api.uplinkswitch.com/v1/chat/completions. The best Janitor AI model for you comes down to Max tokens and price per message. roleplay-uncensored answers with no thinking step, so it fits a low Max tokens setting. For a build that thinks first, such as glm-5.3-abliterated, set Max tokens to 1,000 or more.

Why does my roleplay model reply with nothing? A model that thinks before it answers spends part of max tokens thinking. At a low setting such as 300 it can run out before the reply starts. Set max tokens (SillyTavern's Max Response Length, Janitor's Max tokens) to 1,000 or more for GLM-5.3 Abliterated and DeepSeek V4 Pro Uncensored. Roleplay Uncensored has no thinking step.

Get a key: prepaid from $10, no subscription, 30-day price lock.