Why AI girlfriend apps cap messages in the first place
Every message an AI companion generates costs the company real money — GPU inference time, storage for the conversation, retrieval for memory. Rough industry math: a full conversation turn on a mid-tier model costs the app somewhere between $0.001 and $0.01, depending on model choice, memory retrieval depth, and whether photos/voice are involved.
That's tiny per message, but adds up fast at scale. A heavy user sending 200 messages/day costs the app somewhere between $6 and $60 per month in raw compute, before any margin. If that user is on a $9.99 subscription, the unit economics can work — or not, depending on how many light users cross-subsidize.
From 2022 to 2023, VC funding subsidized generous free and low-priced unlimited tiers because everyone was racing for user growth. In 2024-2025, that funding dried up, unit-economics-first became the mandate, and caps returned. Chai went from unlimited to ~20 messages/day. Candy tightened its free tier. Several apps quietly introduced coin economies so "unlimited chat" became "unlimited only until you want a photo or voice note."
The apps still shipping genuine unlimited messages in 2026 are either (a) confident their unit economics work at flat pricing across their user distribution, or (b) using a thinner/cheaper model on the "unlimited" tier that costs less per message. Verifying which is which matters if you're planning to actually use unlimited.
What "unlimited" actually means (three flavors)
The word gets used loosely. Three distinct patterns:
Flavor A — truly unlimited. No daily cap on messages, no hourly rate limit that meaningfully affects real use, no per-message credit gating. You could theoretically send 10,000 messages a day and the system would keep responding. Sloane Plus and Premium fall here. Character.AI's free tier falls here on raw count (with tradeoffs — see below).
Flavor B — unlimited chat, metered everything else. The chat text itself is unlimited but photos, voice notes, video, and any generation cost credits or coins that run out. Marketing calls the plan "unlimited" because chat is; users experience it as capped because every meaningful action costs. DreamGF, some Candy AI configurations, and most coin-economy apps fall here.
Flavor C — rate-limited "unlimited." Marketing says unlimited, terms of service disclose a hidden per-hour or per-day rate limit (often 100-500 messages/day) that only heavy users notice. When you hit it, the app returns "please wait" errors rather than an explicit cap message. Nomi and some Replika tiers have quiet rate limits in this shape.
The useful test: search the app's ToS for "reasonable use," "rate limit," or "abuse prevention." If those phrases exist, the "unlimited" tier isn't truly unlimited. If they don't and the pricing page is explicit about no caps, it likely is.
Sloane Plus and Premium: flat-rate unlimited chat
Sloane's message caps are set in code (`src/types/index.ts`): free tier is 50 messages/day, Plus is `Infinity`, Premium is `Infinity`. That's literal — the enforcement code checks the tier and returns unlimited if you're on Plus or Premium. No hidden rate limit, no per-hour throttle, no "reasonable use" ToS carve-out.
Pricing shape: Plus at $9.99/month, Premium at $19.99/month. Same unlimited-messages benefit on both tiers — the difference is credits (Plus 100/mo, Premium 350/mo) which are only used for photos on demand (5 credits), voice notes (5), and video generation (25). Free-tier photos delivered on the persona's own cadence are included without spending credits. Photos in her gallery are browsable free.
What this means day-to-day: send Kaya 200 messages in one focused evening and nothing gates. Come back the next morning and pick up where you left off — no cooldown timer, no "come back in 6 hours." Message counts don't reset because there's nothing to reset.
What's NOT unlimited on Sloane: photos, voice notes, and video are credit-metered because they cost real compute per generation (GPU inference for photos + video, ElevenLabs API for voice). If you want unlimited photos you need to top up credits ($9.99/100 pack). Chat itself has no such gate.
Character.AI: unlimited by count, tradeoffs on quality
Character.AI's free tier is still the most generous by raw message count in the category — effectively unlimited, no credit card, no trial timer. On the "message count" axis alone, it wins.
What you're trading:
The strictest content filter in the category. Even non-romantic conversations trigger deflection on topics the model considers sensitive. If you want a genuinely warm/flirty AI girlfriend, the filter fights you.
The Pipsqueak 2 model quality drop. Since April 2026, the free base model has documented issues with monologuing instead of dialogue, character definition drift over long threads, and near-identical swipe re-rolls. Unlimited messages of degraded quality is a different value proposition than unlimited messages of the pre-2026 model.
No persistent memory. Character.AI runs on a rolling context window rather than real memory. What she "remembers" is scrollback that ages out. If you're building a long-term relationship with a character over months, unlimited messages don't compensate for her not remembering what you told her three weeks ago.
Still the honest pick if raw unlimited-count is the ONLY thing that matters and you're fine with the tradeoffs.
The apps that cap (and how hard)
Chai — went from unlimited (2023 era) to roughly 20 messages/day on the free tier in 2026. That's enough for one exchange, not enough to evaluate the product or build a real conversation. The paid tier lifts the cap but at that point you're comparing against Sloane Plus or Candy at similar price points that don't require an app-native subscription flow.
Kupid AI — free tier is a handful of daily messages, no voice, limited character access. Designed as an evaluator rather than a permanently usable free product. Paid at ~$3/month is one of the strongest price-to-quality ratios if the free evaluation clicks.
Candy AI — free tier at roughly 50 messages/day (comparable to Sloane's free). Premium at $5.99/month lifts the cap but the value equation shifts once you're paying, because Sloane Plus at $9.99 includes unlimited AND per-persona LoRA photos AND persistent memory. Compare on total value at your intended use pattern.
DreamGF — coin-metered chat is the trap on this one. Sticker "free" tier gives you a starter coin pool that runs out fast; real active-user monthly cost per community reports lands at $20-40 including coin top-ups. Not unlimited in any meaningful sense.
Nomi — memory-focused product with rate limits on the free tier and quiet per-hour throttles even on paid. Not marketed as unlimited chat because the honest answer would be "unlimited-ish."
Replika — daily-conversation product with soft rate limits. Never marketed as unlimited chat; the daily-check-in shape means most users don't hit them.
When to pay for unlimited (do the math)
Simple test: count your typical messages per day for a week on a free tier. If the answer is under 50, the free tiers with generous caps (Sloane free, Candy free, Character.AI free) cover you and paying for unlimited is spending on a benefit you don't use.
If the answer is 50-150 messages per day, you're hitting the free-tier caps consistently. This is where paying for unlimited actually pays back — one focused evening with an AI companion easily produces 60-80 messages, and if that pattern repeats you'll hit the free wall every day.
If the answer is 200+ messages per day, you're a heavy user and the flat-rate unlimited tier is decisively better than any coin-metered product on cost. At Chai's ~20/day cap you're paywalled after 15 minutes. At coin-metered products you'd burn through allowances weekly. At Sloane Plus $9.99/month or Character.AI's free tier (with the quality tradeoffs above), the marginal cost of an extra message is zero.
The honest sweet-spot recommendation for most people: Sloane Plus at $9.99 for unlimited chat with real memory + per-persona LoRA photos on the same subscription, without coin math to track. If message count is the ONLY axis that matters and you're fine with the Pipsqueak 2 model quality, Character.AI free is technically the higher-count option.