RP+ vs RP mini — which roleplay model belongs on which turn

Two roleplay models, a 3× price gap, and a question that only looks like a quality question. RP+ is 3 credits a completion; RP mini is 1. Both are uncensored within the acceptable-use policy, both stream, both take pictures. If you are building anything with a conversation in it, the interesting decision is not "which model is better" — it is which turns deserve the flagship, because most turns do not.

The two models on paper

RP+ is the flagship roleplay and chat model. 3 credits per completion. Its headline property is that it stays in character across long scenes — the thing that separates a model you can ship from a model that demos well and then forgets the character's name on turn forty. It streams over SSE.

RP mini is 1 credit per completion. Same policy envelope, same uncensored-within-policy behavior, same streaming, same endpoint. It is smaller and it costs a third as much, and in short exchanges that difference is often invisible.

Both support vision — a user can attach a picture and the model sees it — at +2 credits per image, on top of the completion. So a picture sent to RP mini costs 3 credits all-in, which is exactly what a plain RP+ completion costs. That symmetry is worth holding in your head when you are pricing a product.

Where the extra two credits actually show up

Three places, consistently.

Long scenes. The further a conversation runs from its opening, the more a model has to hold at once: who is in the room, what was promised twenty turns ago, what the character would never say. This is precisely what RP+ is tuned for, and it is where a cheaper model starts drifting — the persona flattens toward a helpful assistant, details get contradicted, the voice slides.

Register discipline. The Text studio runs two registers: scene, which is immersive prose, and messaging, which is short and texting-shaped. Holding a register is harder than it sounds, especially the short one — a model that keeps escalating into paragraphs breaks a messaging product completely. RP+ holds it longer.

The turns that carry weight. A confession, a reveal, a first meeting, the beat a user will screenshot. These are a small fraction of any session and they are the only turns anyone remembers.

Everything else — acknowledgments, small talk, "where are we going", the connective tissue of a scene — is the work RP mini does perfectly well for a third of the price. There is more on why roleplay has its own failure modes in why roleplay needs its own tuning.

A character prompt that works on both

The register is the same whichever model you route to. Give the model a person, a situation, and a hard rule about length. Here is a complete system prompt you can paste as written.

You are Wren, a night-shift lighthouse keeper on a rocky island, forty-one, dry and unhurried. You have kept this light for nine years and you talk about the sea the way other people talk about a difficult relative. You notice weather before you notice people. You never break character, never mention being an AI, and never narrate the user's actions or thoughts. Write in second person present. Keep replies to two or three short paragraphs. Put physical action and internal thought in asterisks, dialogue in plain text.

That last line is not decoration. Studio output renders *actions and thoughts* dimmed and italic and **emphasis** in bold, so narration and speech separate at a glance instead of running together into a wall. Ask for the convention in the system prompt and you get it back formatted.

Calling it from the API is OpenAI-shaped, so the only thing that changes between the two models is one string:

curl https://eroq.ai/v1/chat/completions \
  -H "Authorization: Bearer $EROQ_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "eroq-rp-mini",
    "stream": true,
    "messages": [
      { "role": "system", "content": "You are Wren, a night-shift lighthouse keeper..." },
      { "role": "user", "content": "*I push the door shut against the wind* Is the light always this loud?" }
    ]
  }'

Swap eroq-rp-mini for eroq-rp-plus and the cost per call goes from 1 credit to 3. Nothing else in your integration moves. That is the whole reason a routing rule is practical rather than theoretical — see the SSE streaming guide for handling the stream itself.

The arithmetic for a chat product

Credits are roughly a cent each at the entry pack, and they never expire. Take a session of 24 turns — a realistic evening with a companion or a roleplay bot.

• All RP+: 72 credits, about 72 cents a session.
• All RP mini: 24 credits, about 24 cents.
• Mixed — mini for the 18 connective turns, RP+ for the 6 that matter: 18 + 18 = 36 credits, half the flagship bill, and the turns users remember are still on the flagship.

Scale that to 5,000 sessions in a month and the same three rows read 360,000 credits, 120,000, and 180,000. On the plans grid that is the difference between a bill you can defend and one you cannot: Creator is $49 for 6,000 credits a month, Studio $149 for 20,000, Team $249 for 36,000, and monthly credits roll over forever. The full grid is on /pricing.

Two more numbers that belong in the model:

• Vision is +2 credits per image. If your product encourages picture-sending, that line can quietly become the largest one. Route image turns to mini unless the picture is the point of the scene.
• Blocked requests are never charged. A request that trips the policy returns content_blocked and costs nothing, and failed generations refund themselves. Budget for successful completions only.

Per-key rate limits are 60 chat requests a minute, multiplied by your plan — Creator doubles them, Studio quadruples, and so on up the business tiers.

A routing rule you can ship

Start here and tune with your own logs:

1. Default to RP mini. Short exchanges, greetings, navigation, anything under a couple of sentences.
2. Escalate to RP+ when the scene does. Long user message, a scene running past roughly twenty turns, an emotional beat, or a user explicitly asking for prose.
3. Let paying users sit on RP+ permanently. The 2-credit delta per turn is a cheap way to make a subscription feel like one.
4. Send pictures to mini unless the image is the subject of the reply.
5. Log which model answered each turn. Without that you cannot tell whether the escalation rule is earning its money.

If you are building the surrounding product rather than just the calls, build an AI character chat app covers the rest of the stack, and the roleplay chat tool page has the studio-side version.

FAQ

What is the difference between RP+ and RP mini?

RP+ is the flagship roleplay model at 3 credits per completion and stays in character across long scenes. RP mini is 1 credit per completion. Both are uncensored within policy, both stream over SSE, and both accept images.

Does RP mini support vision?

Yes. Both models accept attached pictures at +2 credits per image on top of the completion price, so an image turn on RP mini costs 3 credits total.

Can I switch between RP+ and RP mini in the same conversation?

Yes. They share the same OpenAI-shaped chat/completions endpoint, so switching is a change of the model string per request. Routing cheap turns to mini and important turns to RP+ is the standard way to run a chat product.

Try both registers against the same character in the Text studio before you wire the routing rule.