Best models for JanitorAI: what actually holds up per message

Quick answer: the lowest-priced models that still hold a JanitorAI chat

DeepSeek V4 Flash (deepseek-v4-flash) and MiMo V2.5 (mimo-v2.5) sit at the floor of the usable price range, both priced at roughly $0.126 per million input tokens and $0.252 per million output tokens. At typical JanitorAI message sizes, that lands well under a cent per message, card and history included. MiniMax M2.7 (MiniMax-M2.7) is a modest step up for slightly better recall in longer chats, and DeepSeek V4 Pro and MiMo V2.5 Pro roughly triple the floor-tier rate for meaningfully steadier character consistency. The lowest sticker price is not always the lowest overall cost. A model that drifts out of character or cuts a reply short forces a regenerate, and a regenerate bills as a second full message, card and all. This page walks through actual per-message math for the realistic budget tier, a comparison of where each model actually earns its price, and the exact JanitorAI proxy fields to paste in once you have picked one. This is written for anyone bouncing between two or three JanitorAI configurations trying to find the one that stretches a small monthly budget the furthest, not for anyone chasing the single lowest number on a spec sheet. The two goals usually point at the same handful of models anyway, since the floor-tier price band above is wide enough to cover most casual to moderate chatting without upgrading to a Pro-tier or flagship model at all.

Why budget models can still add up: where JanitorAI sends your tokens

None of this changes because the model is low-priced. A $0.126-per-million model still resends the same card and history as a $4-per-million one; the per-token rate is just lower, so the same volume costs less. The next section runs the actual numbers so the budget label has a dollar figure attached instead of just being relative.

• Character card and persona: resent in full with every single message, no matter how low the model is priced.
• Chat history: JanitorAI keeps recent turns in the context it sends, so a forty-message conversation is resending most of those forty messages by message forty-one.
• Any custom prompt or extra system instruction you have added: it rides along on every call the same as the card does.
• Output length: JanitorAI's max response tokens setting caps what a single reply can cost, but setting it too low causes cut-off replies, and a cut-off reply usually gets regenerated, which bills the message twice.

Worked example: a 1,000-message month, card and all

Assume a typical detailed character card plus recent history: about 5,000 input tokens and 300 output tokens per message, which is a realistic mid-size setup rather than a bare-bones one. The table prices that message shape across the budget tier for a 1,000-message month, a heavy month for most users but a reasonable one for anyone chatting daily. These are APIsRouter catalog rates, an OpenAI-compatible gateway with pay-as-you-go billing: no subscription, global models priced 20 percent below official list, Chinese models priced below their official rates, and a first top-up that adds a 100 percent balance bonus. A few dollars at these rates covers a genuinely heavy month of chatting rather than a single afternoon.

Where the bill quietly grows past the sticker price

None of these are reasons to avoid the budget tier; they are reasons to know which habits move the needle before picking a model. The comparison below is about where each budget model actually holds up, not just its list price.

• Group chats with multiple personas: each active character resends its own card every turn, so a three-character group chat can roughly triple the per-message input cost of a one-on-one chat.
• Regenerates: hitting regenerate because a reply cut off or broke character bills a full second message, card and history included, not just the new output.
• Long-running single chats: JanitorAI keeps appending to history rather than summarizing it, so message five hundred in one continuous chat resends more than message five did, even on the identical model.
• Output pricing: every model in this comparison prices output at two to five times its input rate, so a character that tends to write long replies can push a budget model's bill higher than its input price alone suggests.
• Free-tier throttling: built-in or free-tier options often queue or rate-limit at peak hours, which is what pushes steady chatters toward a low-priced paid model in the first place.

Budget-tier models compared: where each one earns its price

All seven models below are priced for volume, but they are not interchangeable. Context window, how well a model tracks a card over a long chat, and how it handles a crowded group scene all vary even within a tight price band. Treat the table as a starting shortlist rather than a ranking: the right pick depends more on how long your typical chat runs and how many characters are active in it than on which row has the lowest number in the price column.

Six ways to keep the bill near the budget-tier number

None of these require switching models to save money; most of the gap between a $0.71 month and a surprise bill comes from habits, not from which row of the table you picked. Fix the habits first, then decide whether the floor tier is enough or a Pro-tier model is worth the extra couple of dollars a month for your specific chats.

• Trim the character card to what the model actually needs. A tight 300-word card does the same job on every message as a rambling 1,500-word one, at a fraction of the tokens.
• Set a sane max response length in JanitorAI's generation settings so replies don't cut off mid-sentence and force a regenerate that bills the whole message twice.
• Start a fresh chat instead of scrolling one conversation past a few hundred messages; JanitorAI resends more history the longer a single continuous chat runs.
• Test a new model with curl before wiring it into JanitorAI, so a bad model ID or a stale key doesn't burn a wasted attempt inside the app itself.
• Keep group chats limited to characters you are actually talking to; an idle persona sitting in a group scene still resends its card on every turn.
• If you are comparing two or three budget models side by side, route them through a single OpenAI-compatible endpoint instead of juggling separate accounts and separate top-ups for each provider.

Config: point JanitorAI at any budget-tier model

The proxy fields stay identical no matter which model you pick; only the model name changes. The example below points JanitorAI at APIsRouter with deepseek-v4-flash; swap the model field for mimo-v2.5, MiniMax-M2.7, glm-5, or kimi-k2.6 to test any other row from the comparison table above.

Configuration name:  budget-tier-deepseek
Model name:          deepseek-v4-flash
Proxy URL:           https://api.apisrouter.com/v1/chat/completions
API key:             sk-...
Custom prompt:       (optional, leave blank to use the character card as written)

FAQ