The best API for Janitor AI, ranked for 2026

Quick answer: which API is best for Janitor AI

If you just want a model name to paste into Janitor AI tonight, take one of the four picks above and skip to the setup section. DeepSeek V4 Flash is the default recommendation because roleplay burns tokens fast, and Flash keeps a heavy chat habit under five dollars a month while still writing in character. The rest of this page shows the full ranking of eight models, the exact math behind the monthly cost estimates, what each provider's content policy means for fiction, and the config values Janitor AI actually needs.

• Best overall: DeepSeek V4 Flash (deepseek-v4-flash). The community favorite for roleplay. Expressive character writing, holds long chat histories, and costs a few dollars a month.
• Best prose on a budget: DeepSeek V4 Pro (deepseek-v4-pro). Same family, noticeably richer descriptions and better long-scene coherence, still cheap.
• Best for mature fictional themes: Grok 4.5 (grok-4.5). xAI's published usage policy explicitly permits mature themes in fictional contexts for adult users.
• Best pure writing quality: Claude Sonnet 4.6 (claude-sonnet-4-6), for SFW creative roleplay only. Anthropic's policy prohibits explicit content.

How the ranking works

No single model wins all four criteria. The ranking below optimizes for the typical Janitor AI user: long daily chats, a fixed hobby budget, and zero patience for refusals or timeouts in the middle of a scene.

• Cost per RP session. Janitor AI resends your whole chat history and character card with every message, so input token pricing dominates. A model that looks cheap on output rates can still be expensive in practice.
• Context length. Long chats plus detailed character definitions need a model that stays coherent deep into a conversation instead of forgetting earlier scenes.
• Policy fit. Each provider publishes a content policy, and they differ a lot on fictional content. Picking a model whose policy matches your stories avoids refusals mid-scene.
• Stability. Roleplay traffic peaks in the evening. Models were weighted by how reliably they respond under load: timeouts, rate limits, and error rates all break immersion.

The rankings

The DeepSeek V4 family takes the top two spots because it is the rare case where the cheapest option is also the community favorite: approximate community sentiment consistently puts it first for character voice and willingness to follow fictional scenarios. MiMo and GLM-5 are strong value alternates with a different prose flavor; GLM-5 in particular writes vivid scene description. Kimi K2.6 earns its spot on long-history coherence and is worth testing if your chats run for hundreds of messages. Gemini 3.5 Flash is fast and very stable but applies stricter automated safety filtering, which makes it a better fit for lighter, SFW-leaning chats. Grok 4.5 is the pick when your fiction runs into mature themes, since xAI's published policy addresses that use directly for adults. Claude Sonnet 4.6 ranks last only because of cost and its SFW-only policy fit; on raw prose quality it beats everything else on this list.

What each model costs per month at real usage

Approximate community-reported numbers for a regular Janitor AI user land around 23,000 tokens per request and about 1,200 requests per month. The estimates below assume that request is split roughly 22,000 input tokens (character card plus resent history) and 1,000 output tokens (the reply), which works out to about 26.4M input and 1.2M output tokens per month. The rates below are catalog prices on APIsRouter, an OpenAI-compatible pay-as-you-go gateway that carries every model in this ranking on one key with no subscription: global models run 20% below official list prices, Chinese models below their official rates, and a first top-up adds a matching +100% balance. One note on Grok: grok-4.5 is xAI's flagship general model, token-billed at $1.60 input and $4.80 output per 1M tokens. Those are flagship rates, so at 23K-token requests it lands well above the budget tier, but its 500K context window means even very long chats stay inside one window.

Understanding content policies before you pick

Every provider publishes a usage policy, and for roleplay the differences matter more than benchmarks. This is neutral ground: knowing where each policy stands saves you from refusals mid-scene and from violating terms you never read. The DeepSeek V4 family, MiMo, GLM-5, and Kimi K2.6 have comparatively flexible policies for fictional content, which is why they dominate roleplay communities. Grok 4.5 is the most explicit about it: xAI's published policy permits mature themes when they appear in clearly fictional contexts and the user is an adult. Gemini 3.5 Flash applies stricter automated filtering, so treat it as a fast option for lighter stories rather than a general roleplay workhorse. Claude Sonnet 4.6 deserves a special note. It produces the best character prose of anything listed here, but Anthropic's policy prohibits explicit content, so use it for SFW creative writing and expect refusals if a scene drifts past that line. Finally, whatever model you pick, Janitor AI's own terms apply on top of the provider's policy, including its age requirements. Pick the model whose written policy matches what you actually write, not the one you hope will look the other way.

Set it up in five minutes

Janitor AI takes any OpenAI-compatible API through its proxy configuration. The important detail that trips most people: Janitor AI wants the full chat completions endpoint, not just a base URL. Get an API key first (on the gateway from the pricing section, you pay at /topup and the key arrives by email, no account signup), then open a character chat, go to API settings, choose the proxy option, and enter these three values:

API / Proxy URL:  https://api.apisrouter.com/v1/chat/completions
Model name:       deepseek-v4-flash
API key:          sk-APIsRouter-...

Notes for SillyTavern users

If you run SillyTavern alongside Janitor AI, the same key works but the URL format differs. SillyTavern wants the base URL ending at /v1 (https://api.apisrouter.com/v1); once you paste it with your key, the model list auto-populates from /v1/models and you pick models from a dropdown instead of typing IDs. That difference explains most "invalid URL" errors people hit when switching between the two frontends: Janitor AI needs the full /chat/completions path, SillyTavern needs the path cut at /v1. Keep both forms in a note and swapping frontends takes seconds. Swapping models is also trivial from either side, so a common pattern is running deepseek-v4-flash for everyday chats and switching the model name to deepseek-v4-pro or glm-5 for scenes where you want heavier prose.

FAQ