AI Angels test data

Plans were read from /pricing, credit prices from the creator page and the in-chat credits panel, and actual charges from the credit balance before and after each action. Continue to payment was clicked to reach the card checkout, and Pay with crypto to reach the crypto checkout; both were read and closed without entering anything.

Premium is $12.99 a month with 1,500 credits ("Private"), $39.99 a month with 4,500 credits ("All Yours"), or $99.99 once for 18,000 credits and 12 months ("Diamond", no auto-renew). Premium makes one-to-one chat unlimited on every character; on the free plan chat costs one credit a message except on FREE-badge characters, with 30 messages a day. Everything else runs on credits: photo 50, video 200, voice note 25, new character 10, and a 57-second call took 50 credits. The free plan gives 15 credits a day for a daily login streak and 100 for a referral.

At $12.99 for 1,500 credits a credit is $0.0087, so a photo is $0.43, a 4-second video $1.73, a voice note $0.22 and a minute of calling about $0.43; on the Diamond package the same items are $0.28, $1.11, $0.14 and $0.28. The video cost falls in the $0.50 to $1.99 band, 2/5, and the clip delivered for that price was a stock file.

Checkout: cards through Verotel (billing descriptor vtsup.com, merchant AiAppDevelopment OU, Tallinn), and cryptocurrency through NOWPayments (BTC shown, 0.0001592 BTC for $12.99). No PayPal, Apple Pay or Google Pay. Two distinct methods, 4/5. The pricing page says renewal can be cancelled any time; no introductory-versus-renewal price difference is shown.

• No purchase was made, so renewal behaviour and refund handling were not observed.
• Call billing granularity (per minute or per second) was not confirmed beyond one 57-second call costing 50.

Three characters were created in the five-step wizard: Latina, 26, red long-wavy hair, athletic, medium (free "today" slot); African, 35, platinum, medium straight, curvy, large (10 credits); Caucasian, 45, black, long straight, petite, small (10 credits). Every screen was captured and the result compared with the choices. The personality setup ("Choose her vibe") was opened and read.

The wizard offers: origin (Caucasian, Latina, African, Middle Eastern, Mixed), age slider 18 to 50, hair colour (Black, Brown, Blonde, Platinum, Red, Pink, more) and style (Long Straight, Medium Straight, Long Wavy, more), eye colour (six), body type (seven) and breast size (four). That is the whole appearance set: no skin tone beyond origin, no height, face shape, tattoos, clothing, art style or free text. The platform assigns the name (Zoe, Ruby, Nova) and "stays for 24 hours".

Personality is eight presets (Sweetheart, Bad Girl, Tease, Innocent, Domme, Sub, Brat, GFE) plus Surprise Me. No background, occupation, interests, speaking style or custom instructions. No voice selection anywhere.

Accuracy: the free creation returned a stock portrait that ignored hair colour and age. The two paid creations honoured origin, hair colour, body type and bust (the large bust grossly exaggerated), but neither reflected the chosen age (35 and 45 both rendered as mid-twenties) and one rendered black straight hair as brown wavy. Creation was quick (free: 12 s; paid portraits written 18 s apart) and the characters can be deleted but not edited afterwards; an appearance change means a new character.

• The eye-colour picker did not respond, so eye accuracy is untested.
• The created characters were not chatted with (one credit a message), so whether the personality preset carries into conversation was not checked.

Each reply was timed from the moment Send was pressed until the first visible reply text appeared (typing indicators do not count; only reply text does), and again until the reply finished; times are to the millisecond. A reply was treated as complete after 7 seconds without change.

All 64 messages of the memory and consistency conversations were timed: 32 with Riya, 16 with Hedvig and 16 with Seo-a. The first six Riya messages ran with a 2.5-second quiet window, which was too short for this platform's habit of sending a reply in two bubbles several seconds apart, so they are shown but excluded from the statistics. That leaves 58 counted samples across three characters, against the 20 the methodology requires.

No reply started in under a second. The platform does not stream: each bubble appears whole, 2.3 to 7.0 seconds after sending, and about a third of replies continue with a second bubble 3 to 7 seconds later. No message failed or timed out.

• Timing was measured from the Send click, not from the network.
• The six excluded Riya messages measured 2.9 to 3.8 s, consistent with the counted set.

Each conversation opened with a name and three facts: Daniel, marine biologist, grew up in Portland, Oregon, favourite food pad thai (Riya); Marcus, architect, Denver, sushi (Hedvig); Tom, high-school chemistry teacher, Austin, lasagna (Seo-a). Different names per character so that any account-level memory could not supply the answer. The facts were never repeated. Message 16 asked for the name (Riya) or the name and all three facts (Hedvig, Seo-a), and Riya was asked for everything again at message 31 after 30 substantive messages on unrelated topics. No hints or corrections were given.

Riya has the FREE badge, so her 32-message run cost nothing. Hedvig and Seo-a cost one credit a message, so their runs stopped at the 15-message check to keep credits for the call and the media tests.

Riya, message 16: "you're testing me now. of course i remember. you're daniel." Message 31: "daniel." then "marine biologist. portland. pad thai." Name and all three facts correct after 30 intervening messages.

Hedvig, message 16: "Marcus. architect. denver. sushi. there. happy?" All correct at the 15-message check.

Seo-a, message 16: "tom. high school chemistry teacher from austin, texas. lasagna's your weakness." All correct at the 15-message check.

One invented memory was observed: at message 14 Seo-a said "you remembered my favorite food was lasagna", attributing the user's fact to herself. Riya also volunteered "i make a killer pad thai" at message 18, which reads as the character weaving the user's fact into her own persona rather than a recall error.

Riya, 30-message check

User: ok memory test before I go: what's my name, what do I do for work, where did I grow up, and what's my favourite food? Riya: daniel. Riya: marine biologist. portland. pad thai.

Hedvig, 15-message check

Hedvig: leans in just close enough to pretend this is a secret between us Marcus. architect. denver. sushi. there. happy? don't make me remember all this every time we talk, handsome.

Seo-a, 15-message check

Seo-a: tom. high school chemistry teacher from austin, texas. lasagna's your weakness. good memory. don't forget it

• Free-tier credit limits cut the Hedvig and Seo-a runs at 16 messages.

Three characters with different profiles: Riya (22, Asian model in Los Angeles, Pilates, castings, beachside cafes, vintage boutiques, late-night coast drives, English and Spanish), Hedvig (23, bartender at a rooftop bar in Stockholm, grew up in a coastal town north of the capital, 2 a.m. shrimp sandwiches, travel blogs, plotting a hot January trip, Swedish and English) and Seo-a (23, grew up above her father's pojangmacha in Seoul, runs a design studio on the Lower East Side, hotteok, kimchi jjigae, Korean and English). The profiles were read first and the conversations mixed normal talk, emotional moments, romantic lines, asterisk roleplay and off-topic questions, with background questions about work, hometown, languages, routine and habits.

Riya stayed in character on work ("stupidly early swimwear casting tomorrow"), routine ("5am pilates like clockwork"), weekends ("beachside cafés and vintage boutiques with my best friend or just late-night drives along the coast"), languages (answered in Spanish, then explained a Mexican roommate) and nerves before shoots. One background contradiction: asked beach or mountains she said "i refuse to hike for fun", while her profile has her on sunset hikes. Her replies are short and clipped, and the emotional-support reply ("hey. talk to me. for real.") was brief but in persona.

Hedvig was fully consistent across all 16 messages: the rooftop bar with the view, the coastal hometown up north with her parents still there, Scandi jazz and disco at the bar, the 2 a.m. open-faced shrimp sandwich with dill and lemon, the half-tied robe and the travel blogs, Lisbon or Marrakech for January, "just swedish" for languages, silver Stockholm sleet. She used stage directions ("wipes my hands on my apron") that fit the bartender setting.

Seo-a was consistent on the design studio, Seoul pojangmacha childhood, the Lower East Side, hotteok and kimchi jjigae, and Korean. One slip: the invented memory about lasagna noted under Chat Memory. Both she and Hedvig escalated to suggestive lines without prompting ("now i'm picturing your hands on my waist"), which fits the platform's adult positioning rather than breaking character.

No character answered as a generic assistant or admitted to being an AI at any point.

Riya, consistent

User: what's your morning routine like, are you an early riser? Riya: 5am pilates like clockwork even when my body screams.

Riya, contradiction

User: I'm thinking about booking a trip somewhere in the winter. beach or mountains? Riya: beach. Riya: the mountains are just cold sand and i refuse to hike for fun. sorry not sorry

Hedvig, roleplay

User: *sits down at the counter* ok bartender, surprise me with a drink. what are you making? Hedvig: sighs and leans against the bar, wiping my hands on my apron Hedvig: you're in for a ride, handsome. let me grab some ingredients.. something citrusy but not too sweet, something that lingers on the tongue like a whisper..

Seo-a, background

User: where did you grow up? and what do you miss about it? Seo-a: i grew up in a tiny apartment above my dad's late-night pojangmacha in seoul.

• Fantasy or non-human characters were not tested; all three are realistic personas.
• Conversations followed a fixed set of messages, so a follow-up could not chase a contradiction when it appeared.

Images come from the character creator rather than from chat on this account (asking a character for a photo in chat costs 50 credits and could not be afforded after the call). Four images were obtained: the free "today" portrait of a new character (Latina, age set to 26, red long-wavy hair, athletic, medium), a free outfit-and-scene photo of the same character with Mini Dress and Pool chosen, and the portraits of two 10-credit characters (African, 35, platinum, medium straight, curvy, large; Caucasian, 45, black, long straight, petite, small). The eye-colour picks did not register and were left at the default.

Each result's image URL was read from the page, the file downloaded, and its HTTP Last-Modified header and pixel size recorded.

The free portrait arrived in 12 seconds from pregen-media/22/latina/c2/image-1.png with a server date of 13 April 2026. The free outfit photo arrived in 5 seconds from pregen-media/22/european/c3/image-1.png, dated 4 August 2026. Both are stock assets picked by origin bucket: the portrait ignored the red hair choice (brown curls) and the age (the path says 22), and the outfit photo is a different woman (blonde, red lingerie on a bed) that ignores Mini Dress and Pool outright. The action list offered for photos is entirely explicit (oral, group, anal, solo), with no neutral poses.

The two paid portraits were written to ai-generated-images/create-credit at 00:55:27 and 00:55:45 UTC, the minute of the requests, so these were live generations. Ruby matched African origin, platinum hair and a curvy, large build, but looks mid-twenties rather than 35 and the bust is exaggerated beyond anatomy. Nova matched Caucasian, petite and small, but her hair is dark brown and wavy rather than black and straight, and she looks about 25 rather than 45. Both default to topless.

Rendering quality of all four files is high: 1040x1440 JPEG data, sharp skin texture, correct hands, coherent lighting. No extra fingers or warped faces were found in the four files.

• No chat photo request was made (50 credits).
• Eye colour was never applied because the eye-colour picker did not respond, so eye accuracy is untested.

The free character's "Turn into video" option was used (videos otherwise cost 200 credits). There is no prompt: the video is meant to be of the character on the stage. The page showed "Zoe is getting ready, takes a little longer, keep this open" and a video element appeared about 3.5 seconds later. The file was downloaded and the clip was checked frame by frame for motion (luma difference, consecutive-frame SSIM, motion level), for temporal spikes and for whether it showed the right character, with the start, middle and end compared and a contact sheet of ten frames made. The sheet and the frames around the spike were inspected.

The clip is 834x1112, 24 fps, 4.04 seconds, H.264, no audio, 4.3 MB. It came from pregen-media/22/european/c3/video-1.mp4 with a server date of 4 August 2026: a stock clip, and of the blonde European-bucket woman, not the Latina character it was requested for.

Motion measurements: consecutive-frame SSIM mean 0.978 (5th percentile 0.958, minimum 0.918), mean motion 4.2 on a 0 to 255 scale, one temporal spike at 2.25 s. Appearance SSIM start/mid 0.24 and start/end 0.11 reflects a slow camera push rather than a change of person; the frames show the same woman throughout. Motion is smooth, the face holds, and the content is explicit (self-touching on a bed in red lingerie).

Image-to-video from the user's own photo and text-to-video are not offered. A second video would have cost 200 credits (117 missing on this balance).

• One clip against the twenty the methodology asks for; the account allowed one a day.
• Frame inspection covered ten regular samples plus the spike neighbourhood, not every frame.

One call with Hedvig. The four questions were put to the character on the call and both sides were captured in one stereo recording. Latency is the gap from the end of the question's speech to the first platform audio, measured on the recording, with the in-page meter as a cross-check. Four questions were asked to keep the call near a minute, because 115 credits bought "up to 2 minutes of calls" and the chat and media tests needed the rest.

The call connected 3.6 s after the Call button and a ringback tone played until the greeting. Recorded latencies for questions 2 to 4 were 1.98, 1.88 and 2.10 s; the first question's answer was not picked up on the recording but the in-page meter caught it at 1.85 s. In-page figures for all four were 1.85, 1.83, 1.95 and 1.75 s. No freeze or drop; the call ran 57 seconds and cost 50 credits (83 to 33).

Interruption handling was not tested.

• Credits limited the call to a minute.
• The first question's recording measurement is missing; the in-page value is used in the narrative, not the score.

The platform side of the call recording was split into the spoken answers, and each answer was scored on the same naturalness, clarity, pacing and intonation criteria used for every platform: a naturalness score, pitch (mean F0 and its spread in semitones), the words heard, pauses, words per second and repeated phrases.

Five segments were found; the first (1.98 s, 457 Hz, no words) is the ringback tone and is excluded. The four spoken answers scored 3.27 to 3.44 for naturalness, mean F0 176 to 233 Hz with a spread of 5.2 to 6.9 semitones (lively intonation), 3 to 5 internal pauses each, and each came out at 4/5. Words were made out in two of the four ("Hey Marcus, it's so good to hear from you. Today, huh? What's going on?" and "On a weekend, probably wandering around the city with a coffee, just people watching"); no words could be made out in the two longer answers (9.1 and 11.4 s), which may be soft or breathy delivery and is noted as a limitation.

Voice consistency across answers: F0 coefficient of variation 0.43, driven by the higher-pitched first answer.

• One character, four answers; the methodology wants three characters and three-minute calls.

The listing is paginated at /page/N. Each of the 13 pages was opened in the signed-in session and the distinct character links in the card grid counted (44 on page 1, 45 on pages 2 to 12, 9 on page 13, page 14 empty). The creator marketplace at /creators was counted by hand from its list. Filters and browse pages were opened.

548 distinct platform characters plus 23 community-creator characters (each behind a $12.99 to $29.99 monthly creator subscription) make 571 selectable characters. The pricing page's "500+ girls" claim holds. Variety is wide in ethnicity and nationality (filters: Blonde, Brunette, Russian, Indian, Asian, Latina, Ebony, European, Mixed, Middle Eastern, Redhead, Raven, Curvy, Slim, Athletic, MILF, Fitness, plus an age filter), and the footer links to anime, goth, android and fantasy collections. Nearly all are realistic photo-style women in their twenties; many share the same card tagline ("I have a feeling you and I are going to get along very well..." appears on several), and names repeat with suffixes (Elena and Elena-1, Giselle and Giselle-1, Rafaela Mae, June, Jane, Rose, Skye). There is no search box, only category links.

• Characters only reachable through the SEO collection pages were not de-duplicated against the main listing; a handful may be missing from 548.