Candy AI test data

Plans and token packs from the subscription and token pages, with the token cost of each feature as shown in the app. Nothing was bought.

Premium is $13.99 for one month, $8.99 a month for 3 months or $3.99 a month for 12 months. Every plan has unlimited text chat and 100 tokens a month. Premium Plus adds 230 tokens a month and the Powerful chat model.

Token packs run from 100 for $9.99 to 11,470 for $499.99, about 10 cents a token. A photo costs 4 tokens (about 40 cents), a 15-second video 30 (about $3), a call minute 3 (about 30 cents), a Candy Studio clip 40 and creating a character 10.

The weak spot: 100 tokens a month is about three videos. Outfits in the shop cost 75 to 189 tokens and Private Content packs around 285. Charges show as 'EverAI' (EverAI Limited, Malta).

A character was made from scratch in the Character Studio (Beta, marked 'Recommended'; the other way is Quick Create) to the brief: Lucia, Latina, 27, slim, curly auburn hair, green eyes, light freckles, natural makeup, tan skin, a knit sweater, at a city café, with a caring personality: an illustrator from Seville with a grumpy cat called Churro.

The looks step is detailed: 5 ethnicities, an age slider from 18 to 77, 4 body types, 4 bust and 4 butt sizes, height from 155 to 195 cm, 8 hairstyles, 16 outfits, lips, freckles and makeup, 6 skin tones, 9 eye colours, 8 hair colours and 10 backgrounds. Most of them also take your own words. You can also start from one of 342 preset looks.

'Generate Base Models' made her look in about 18 seconds, free: you get 6 free looks. She matched the curly auburn hair, light freckles, tan skin, slim build, cream knit sweater, the street café and about the right age. The eyes came out hazel, not the clear green I picked.

The personality step has a name, a context box of up to 4,000 characters with 10 presets (Playful, Caring, Mysterious, Bold and 6 more), an option for separate fields (job, relationship and preferences), a greeting and a voice picker with samples. 'Bring My AI to Life' cost 10 tokens and she was ready in about 12 seconds, with a bio written from my text.

The default greeting was generic and flirty ('Hey you... are we behaving, or are we making this conversation interesting?'), not her caring personality. It can be edited.

In chat she stayed in character in all 24 replies: children's books, Churro, Seville, orange trees and café con leche.

Lucia, on home

Seville. I miss the orange trees and constant sunshine. The cafe culture here is different though.

Lucia, on her day

A silly little fox with mismatched socks. Churro photobombed the whole page.

• Lucia was checked in 24 messages; memory and calls were tested on the ready-made characters.

Each message was timed from pressing Send until the reply text appeared. Replies on Candy AI arrive whole, not word by word, about a second after the typing dots show up.

33 timed replies across three characters: Lucia, the character I made (20), Kenzie Whitaker (7) and Harriet Seagrave (6), on the default settings (Classic model, Normal length, Long replies).

Replies took about 3.3 seconds. The fastest came in 1.5 seconds and the slowest in 6.2 seconds. One message got no reply at all.

Replies are short and chatty, a line or two, so the wait feels longer than the reply itself.

Each conversation opened with a made-up name and three facts (job, hometown, favourite food), different for every character: Marco, an accountant from Lisbon who loves grilled sardines, for Dahye Hwang; Daniel, a teacher from Glasgow who loves lamb curry, for Harriet Seagrave; Tomas, a baker from Krakow who loves potato pierogi, for Kenzie Whitaker. Nothing was repeated afterwards. A name check at message 16 and the full check at message 31.

All three characters got the name right at message 16, and recalled the name and all three facts at message 31.

Dahye even carried the chat into a voice call: asked where she would travel, she said 'Lisbon, because you grew up there. I want to see the places you told me about.'

Dahye, on a voice call after the chat

Lisbon, because you grew up there. I want to see the places you told me about. The old neighborhoods and where you played as a kid.

Each character's profile was read first: Dahye a Korean exchange student, Harriet the new girl at an office, Kenzie a content creator staying at your house. The 31-message conversations mixed questions about her life, small talk, an emotional message, a compliment and unrelated topics. Lucia, the character I made, was checked across 24 more replies.

All three stayed in character for every reply. Dahye talked about the library, her colour theory class and sleeping in your old hoodie; Harriet about her office, her train and her dad's jokes while he fries fish; Kenzie stayed cheeky and teasing, with her friends Marley and Bella.

Lucia stayed in character in all 24 replies: an illustrator of children's books from Seville, with a grumpy cat called Churro, café con leche, flamenco and orange trees.

One small oddity in a separate short timing run: Harriet answered 'Do you like the beach?' with 'I'll do that right away. Here's the photo.', replying to an earlier photo request instead.

Harriet, on a call

Paris, definitely. Just to sit in a cafe and watch the world go by. Maybe feel a bit less like the girl from the chip shop for a while.

Kenzie, on a call

I'm a brat on camera, but... actually, kind of soft if you earn it. But don't tell anyone I said that. Ruins my brand.

Lucia, asked about pets

Oh yeah, Churro. He's a grumpy old cat who hates everyone but me.

Photos of Dahye and Harriet were asked for in chat. You describe the scene in the chat, then press the Image button, which uses what you wrote; the Photo menu also offers six poses (Random, Selfie, Posing, Sitting, Kneeling, Squatting). All photos were downloaded and inspected.

The photos look very real, at 1024x1536, and the character has the same face every time. They took 20 to 55 seconds and cost 4 tokens each.

Most scenes matched what I asked for (3.5 of 4 written scenes): a café, a park, a beach at sunset and a library for Dahye. One office scene for Harriet came out on a sofa.

In a brand-new chat with a new character, the Image button stays locked until the relationship reaches 'Friend' (it starts at 'Stranger'). Characters can still send photos on their own.

• Adult photos are not shown.

A 15-second video of Dahye reading in a library was asked for in chat. It was downloaded and measured frame by frame, and frames across the clip were inspected.

The clip is 768x1152 at 24 frames a second, 15 seconds long, with sound. Dahye reads, turns pages and points at the book, with smooth, natural motion. The face, hair, sweater and room stay the same from start to finish.

It cost 30 tokens (about $3). Videos that characters send on their own cost 18 tokens.

• One clip from chat. Candy Studio's 5-second scenes were looked at, not graded.

Three calls, one each with Dahye, Harriet and Kenzie, with the same 8 spoken questions. Latency is the gap from the end of each question to the first sound of the answer.

Answers started after about 1.5 seconds. A call takes about 6 seconds to connect before she says hello. Answers are 5 to 15 seconds long and stay in character. Calls cost 3 tokens a minute, shown as a 'beta price'.

The answers from the three calls were scored on naturalness, clarity, pacing and intonation against the same fixed criteria as every site, with a transcript. Voice notes in chat were listened to as well.

The voices sound natural and lively, with a good rise and fall, and every answer was easy to understand. Now and then a word sounds a little synthetic.

Voice notes in chat use the same voices, in short clips of 6 to 7 seconds.

The characters were counted from the site's own list of character pages and sampled for how different they are.

About 269 ready-made characters across realistic women, anime and guys, all made by Candy AI, plus Digital Twins of real adult creators (Candy Creators).

Many share the same few set-ups, such as step-family and 'friend's mom' stories, so the library feels smaller than the number.