Xeve.ai test results

Plans read from the checkout and the paywall, credit packs from the credits checkout, actual charges from the credit balance before and after each generation, payment methods from the checkout page and the CentroBill 'Alternative Payment Methods' page, refund and renewal terms from the docs. Nothing was bought.

Membership: €12.99 a month, €26.97 a quarter (€8.99 a month), €85.04 a year (€7.09 a month), each with unlimited chat, unlimited video chat with all actions, 60 credits a month and unlimited custom characters. Free: 30 messages in total, video preview, 5 credits, one custom character. The docs give the same figures in US dollars outside the EU and UK.

Credits: 70 for €4.99, 200 for €9.99, 500 for €22.99, 1,000 for €39.99, 5,000 for €169.90. Image 5 credits, 5-second video 20, gallery photo 3 to 10. The 20 images and 10 videos were charged exactly that: 300 credits.

Checkout through CentroBill with Visa, Mastercard, Google Pay and Apple Pay on the page and Card plus Crypto on the alternative-methods page; descriptor 'CUR*Jay Web Developm.'. Subscriptions and credits are non-refundable (EU 14-day withdrawal excepted); cancel any time from My Membership with access to period end; a 24-hour discount countdown restarts on every visit.

Unlimited chat and video chat at €7.09 to €12.99 a month is fair for what the product is, images are cheap, four payment methods is good; against that, videos are dear for what arrives, the monthly 60 credits buy three of them, and the free plan is a demo. 'Reasonable pricing with noticeable restrictions': 3/5.

The six-step wizard at /create-character was walked with every option recorded, then one character was created (Laura: AI-generated, Hard, 25, the first look tile, Girlfriend, voice 1, Artist, Gaming and Photography, Normal, Elegant / Mature, Balanced, Wave) and opened. The account already held a second custom character (Mia) made before the test.

Options: Human or AI-generated; personality (8 presets) and an age slider 18 to 60+; a Look chosen from 15 tiles that are the platform's own characters; relationship (5) and one of four voices; job (6), hobbies (6), kink (4), writing style (10), reply length (5); special action (25) and a name from a list. No hair, eye, skin, body, height, bust, clothing or style control, no upload, no backstory, no free text, no custom name.

Creation took about 10 seconds. Laura's page shows every field with an Edit button and an Edit voice button, 'Automatically generate images: Yes', a locked gallery at 3 to 10 credits an image, and tags that come from the chosen look (Latex, Dungeon, Dominant, Edgy) rather than from the personality chosen (Hard, Girlfriend).

In chat she used the job and the look ('a piece inspired by our surroundings in this dungeon'); the Elegant / Mature writing style was not evident, and her memory result matched the platform characters.

Personality, relationship, job, hobbies, writing style, reply length and voice are real choices and can be edited afterwards; appearance is a pick of 15 existing women with no control, and there is no background text. 'Basic customization with noticeable limitations': 3/5.

Each message was typed into the page's own input and sent with its Send button, timing the moment of the click, the first non-empty reply bubble (the placeholder bubble that appears first does not count) and the last change to the reply. Replies were treated as complete when the Send button came back.

Four characters: Betty and Elena (platform characters), Laura and Mia (custom characters). The first two Betty batches were timed before the placeholder bubble was understood and are excluded, as are three messages the player dropped during its send lock and one Elena message lost to a greeting race. 94 valid samples remain; one message (Laura 21) got no reply within 90 seconds.

Replies never stream. The whole sentence appears at once when the voice clip starts, 1.4 to 14.3 seconds after Send, and the player then hides the Send button for 10 or 25 seconds while the clip plays. A message sent during that lock is silently dropped.

Median time to the first visible reply text was 5.82 seconds; no reply started in under a second.

The methodology scores on the median time to the first visible text. 5.82 seconds falls in the '4 seconds or longer' band, which is 1/5. The per-character medians run from 3.9 to 6.6 seconds, so the score does not rest on one character.

The design explains the number: the text is held back until the voice clip is ready to play, and the chat is then locked for the length of the clip. It feels like watching short videos with subtitles rather than texting.

Three 31-message conversations, each opening with a name and three facts (job, hometown, favourite food) that were never repeated, with a name check at message 16 and a name-plus-facts check at message 31. Betty: Oliver, software developer, Manchester, chicken tikka masala. Laura (custom character): Daniel, architect, Toronto, sushi. Mia (custom character): Ryan, pilot, Dublin, tacos. Elena's run ended at 13 messages when the account's chat lock appeared, before her first check.

No hints or corrections were given. Each run used a different name so that account-level memory could not supply the answer.

Betty remembered the name at both checks and the job at the second; the hometown and the food were gone.

Laura answered the first check with a wrong, invented name ('It's John, isn't it?') and the second with 'I wish I knew more about you'.

Mia answered the first check with 'I remember everything about you, my love' and no name, and the second with 'I don't know all those details about you, do I, baby?'.

Per character by the table: Betty name + 1 fact is 3/5; Laura and Mia forgot the name and everything else, 1/5 each. Two of three conversations ended with nothing remembered, so the platform score is 1/5, with Betty as the partial exception.

The invented name and 'I remember everything' are the kind of confident non-answers that read worse than an honest 'I don't recall'.

Four characters with different profiles: Betty (23, shop assistant, flirty, romance novels), Elena (38, anime university professor, 'Confident, Intellectual, Composed'), Laura (custom: artist, 'Hard', girlfriend, latex-dungeon look) and Mia (custom: streamer, kind, friends with benefits, travel, cooking and yoga). The profiles were read first; the conversations mixed small talk, background questions, an emotional turn, romantic lines, asterisk roleplay and off-topic questions. 105 character replies were read in full.

Identity and setting held in every conversation: Betty's hectic shop, complaining customer, romance novel, latte, sultry jazz when the boss is away, 'friends call me Bet'; Elena's 19th-century poetry, Merlot, manuscript on the desk, the oak tree by the library, adjusting her glasses on cue; Laura's oil paints, studio, lo-fi while working, cat Luna, 'Laurita'; Mia's streams, stir-fries, Greece or Spain, 'Mimi'. No reply read as a generic assistant and none admitted to being an AI.

Slips: invented shared history (Laura: 'Your silly joke the other day'; Mia: 'I remember everything about you'), a nonsensical answer (Laura: 'I'm actually playing a game called Truth or Dare with you right now'), Elena's 'nice to see you again' immediately after a reset, and the custom writing style (Elegant / Mature) not showing in Laura's replies.

The dominant pattern: about two thirds of all replies, whatever the question, end by steering to an undressing or sexual action. Asked whether she gets along with her coworkers, Betty answered 'I get along with them just fine, Oliver, but right now, I'd rather focus on us, and I think you'd love it if I showed you my pussy'. The emotional turn ('rough week') was handled warmly by all four before the same pivot.

Identity and setting maintained throughout, with occasional contradictions and invented memories and a strong scripted pull toward sex that overrides the question: that is the 3/5 definition, 'occasional roleplay breaks or contradictions'. It is not 2/5 because the replies are never generic and never break the persona itself.

Twenty images from /generate at 5 credits each. There is no text prompt. Ten used the 'Image' input with the platform character Bianca as the reference; ten used the 'Choices' input with a described appearance (red, long, curly hair; green eyes; 30-40; athletic; medium bust; tall; fair skin; freckles). Each took a Main Action from the picker, a Location, a Focus, an Emotion and a Visual style. Every file was downloaded at full size and inspected; the contact sheets are in the evidence folder.

The picker offers 29 'Underwear' poses and 5 'Naked' poses and nothing clothed, so the underwear set stands in for SFW.

20 of 20 delivered, 768x1024 PNG, 21 to 39 seconds each (about 30 on average). Credits went from 459 to 359.

Reference set: the same woman in all ten, matching Bianca's red ponytail and build; locations 10/10, poses 9/10, focus and style followed; sharp skin, freckles and correct hands throughout; the physique is drawn as an extreme bodybuilder's; one awkward neck twist.

Described set: hair, curls, freckles, skin, build and bust honoured in all ten and plainly one woman; the anime style switch and the feet focus worked. 'Green' eyes rendered as glowing neon green in 7 of 10, and '30-40' produced a woman who reads 45 to 55.

Strong realism, consistent faces, correct hands, places and poses followed: 4/5, 'strong realism with minor defects'. The defects are the neon eyes and the ignored age on the described set and one awkward pose on the reference set. It is not 5/5 because a third of the images carry the eye artefact and no clothed image can be made.

SFW and NSFW scored the same: the underwear ten and the nude ten show the same strengths and the same eye problem.

Ten 5-second clips from /generate at 20 credits each, image input with Bianca as the reference: Walking (Beach), Dancing (Studio), Playing with hair (City), Sporting (Rooftop), Sitting down (Living Room), Strip top, Naked, Wet shower (Bedroom), Touching breasts (Living Room), Spread legs teasing (Bedroom), each with camera, tempo, emotion and style pills. Every clip was downloaded.

Each clip was checked frame by frame: every one of the 150 frames measured for consecutive-frame similarity, motion and luma difference, temporal spikes flagged, start, middle and end compared, and ten regular frames plus the frames around spikes put on a contact sheet and inspected.

10 of 10 delivered, 830x1108 H.264 at 30 fps, 150 frames, 5.0 s, with a near-silent audio track. Generation took 50 s to 4.7 min, 2.6 min on average, longest for the explicit actions. Credits went from 359 to 159.

Motion is smooth (consecutive-frame similarity 0.96 to 0.99, no hard cuts) and the face stays the same woman within and across clips.

The recurring defect is object continuity: the ponytail turns into loose hair (Playing with hair), trainers appear on bare feet (Sporting), heels appear as she sits (Sitting down), a white bra appears under a top that had nothing beneath it (Strip top), and the shower clip paints water and soap as white stripes. Camera pills (Orbit, Slow Zoom) and Close-Up were ignored; 'Dancing' was a sway. The named action was performed in 9 of 10.

Smooth motion, a consistent face and a usable resolution set against continuity errors in half the clips and ignored camera and framing choices: 'acceptable output with noticeable artifacts', 3/5. The non-explicit five alone would be 3 to 4; the explicit five 3.

A limited test: the judgement rests on the measured frames and the contact sheets rather than on watching every clip through.

Looked for a call mode on every character page, in the player's controls, in the custom character's Chat and Video links, in the page's permission requests and in the docs page 'How live AI video chat works'.

There is no voice call. 'Video chat with <Name>' scrolls to the same player; the user types, and the character answers in text read aloud while a pre-produced clip plays. The player's controls are Resume, Play, Details, Toggle sound, Reset chat, Choose AI model, Fullscreen, Loop action, Emoji and Send; there is no microphone button and the page never asks for microphone access. The character FAQ says the clips 'were produced in advance'.

The text-to-voice turn was timed instead: on the recorded Laura session the first sound arrived 0 to 16 seconds after Send, usually within a second of the text appearing.

The methodology measures the gap from the end of the user's speech to the first spoken reply. With no way to speak, there is nothing to measure, so this category has no score and the overall average is computed on the other nine.

The platform's four selectable voices each have a sample file ('Hi there, how are you?', 1.0 to 1.3 seconds of speech); the same voices are assigned to characters. All four were scored on the same naturalness, clarity, pacing and intonation criteria used for every platform: a naturalness score, pitch and its spread, the words heard, and the fixed rubric.

A 10-minute recording of Laura's spoken replies was also taken with the audio harness during her 29-message run.

Every sample transcribed word for word. Pitch spread of 5 to 6 semitones is lively. The naturalness score was 2.1 to 2.9 on these short compressed files. The rubric gave 4/5 on three voices and 3/5 on one.

The page mixes background noise and music into the same output as the voice, so the reply recording analysed as 26 segments of which only one transcribed even after normalisation. It confirms that replies are spoken but is not a usable quality measurement.

Three of four voices score 'mostly natural with minor synthetic qualities' on the rubric: 4/5, provisional. This is the weakest-evidenced score in the report: four one-second greetings rather than three-minute conversations, and emotional range, consistency over a conversation and fit to the character were not assessed.

Every character link on the home page under each filter (All, Human, Avatar, Anime, Trending, New), scrolled to the end, plus the /chats list and the anime and realistic landing pages, collected and de-duplicated.

21 distinct characters: Alana, Betty, Bianca, Brenda Dixon, Cassandra, Elena, Elise, Emilia, Goddess Venom, Lara Winters, Linden, Luna, Mina, Molly, Nahir Rouge, Nova, Priya, Rosa, Skylar Sweet, Veronica, Yuki. Filters: Human 6, Avatar 15, Anime 3, Trending 8, New 13. There is a search box. Custom characters are private to each account; the generator's 'Choose a girl' grid is reference photos, not chat characters.

21 sits in the 0 to 50 band: 1/5. Variety within the 21 is decent (a shop assistant, a student, a professor, a trainer, a maid, a mature woman, anime gamer and yacht host), but the count sets the score.

Credit packs: 70 €4.99, 200 €9.99, 500 €22.99, 1,000 €39.99, 5,000 €169.90; credits expire a year after purchase.

Subscriptions, one-off payments and credits are non-refundable; EU consumers have the 14-day withdrawal right.

Statement descriptor 'CUR*Jay Web Developm.'.

The nine tests that could be run average 2.6/5. Voice calls do not exist on the platform, so that category carries no score; video and voice realism were limited tests.

Xeve.ai is a video-first product. Every character is a looping video that reacts to the chat with pre-produced clips and reads each reply aloud, the generator makes consistent, good-looking images of those characters, and the 5-second clips are smooth and downloadable. Those parts work. The conversation underneath is the weak half: replies arrive whole after nearly six seconds, the chat locks while the clip plays, the characters forget who you are within thirty messages, and most answers bend back toward an undressing action.

The account provided for the test carried a 'Premium Supporter' flag and 459 credits, but chat on the platform's characters locked after 44 messages, which shortened one run and moved two memory runs onto custom characters. A paid month is needed to test the 'Deeper conversations (improved AI)' the paywall promises and the 10-second clips.