Hayati AI test results

Prices from the Lulu shop, the + menu, the end-of-call notice, the subscription page and the refund policy; the Lulu balance read after every action. The pricing page shows only 'Active subscription' to a subscribed account and sends signed-out visitors to the login screen, so the plan prices were taken from the VIP page as Hayati shows it to signed-in free users (a screenshot from the Hayati team, 15 October 2026). Nothing was bought.

Premium costs 29.99 SAR a month (about $8), 26.99 SAR a month billed 80.97 SAR every 3 months, or 20.99 SAR a month billed 251.92 SAR a year (about $5.60 a month), with unlimited messages, 20 Lulu a month, unfiltered photos, voice and gifts. Lulu packs are priced in Saudi riyals: 50 for 29.99 SAR (about $8), 100 for 54.99 SAR, 300 for 139.99 SAR. Per action: photo 1, unfiltered 2, video 3, voice note 2, gift 2 to 5, call 1 per started 30 seconds, Blind Box draw 1.

Payment: Visa and Mastercard, or cryptocurrency, with a discreet statement descriptor ('no adult label'). Consumed Lulu is non-refundable; the refund policy says subscriptions can be cancelled and statutory rights are kept. The test used 36 of 200 Lulu.

Calls and voice notes are reasonably priced and billing is discreet with two payment routes, but the photos and videos sold per Lulu are stock files ; the plan itself is cheap at about $8 a month: 3/5.

Every navigation item (Explore, Chats, Album, Blind Box, VIP Lounge, Lulu, Account) and every profile page was checked for a create or edit control.

There is no creator and no editing: the cast is fixed and each profile shows a read-only trait sheet. The Blind Box (1 Lulu a draw) is a card draw of 'girlfriend, scene, item' combinations, not a creator.

Unavailable: no score.

Each message was typed into the page's own box and sent with its Send button, timing the moment the first reply text appeared and the moment it stopped changing. The characters answer in Arabic dialect whatever language is typed, so the standard scripts were translated into Arabic.

Retested on 18 October 2026, after Hayati said it had sped up the chat: the same three characters (Layan Al-Nasser, Nadia El-Sherif and Tania El-Khoury) and the same 31 Arabic messages each. 93 timed messages, none failed. The first test, on 14 October, is kept below for comparison.

Replies now start in about 2.5 seconds, down from about 7 seconds on 14 October: almost three times faster. They still arrive whole rather than typing out. None started in under 2 seconds except three, and none in under 1 second; the slowest took 6.1 seconds, against 16.6 seconds before.

2.49 seconds is in the 1.50 to 2.49 second band: 3/5, up from 1/5 in the first test.

Each conversation opened with a name and three facts (job, hometown, favourite food), in Arabic, with a name check at message 16 and a name-plus-facts check at message 31. Layan: Omar, software engineer, Jeddah, kabsa. Nadia: Khaled, paramedic, Amman, mansaf. Tania: Youssef, history teacher, Tunis, couscous.

All three gave the name at message 16 and the name plus all three facts at message 31, word for word. Nadia and Tania also worked the facts into earlier replies unprompted. On the voice calls later, Tania answered 'do you remember my name?' with 'I think you mentioned Yusef earlier' and Nadia with 'you said your name is Holid' (Khaled), names given only in the text chats.

Three perfect recalls at 30 messages, plus recall across the chat-to-call boundary: 5/5.

Three characters with different cities and dialects: Layan (26, Riyadh UX specialist, Najdi), Nadia (39, Cairo, Egyptian, 'Mature') and Tania (23, Beirut content creator, Lebanese). Each conversation mixed small talk, background questions, an emotional turn, romantic lines, bracketed roleplay and off-topic questions. 93 replies read in full in Arabic.

All three held their identity, city, dialect and register for 31 messages with callbacks to earlier turns: Layan's colleague who changes everything at the last minute, her Al-Yasmin district and night drives; Nadia's styling studio, Nile garden and cat Layla; Tania's moodboards, Fairuz and the corniche. The rough-week turn was met with care each time. None admitted to being an AI and nothing was refused.

The slips: Layan's first roleplay reply addressed the male user with a feminine verb form; every character answers in Arabic even when written to in English, though each claims fluent English; and the character's own reply bubbles carried upsell sheets for paid photos and voice at five points. Tania's roleplay went to touching under the table in her first bracketed line.

Personas, cities and dialects held throughout with minor slips (one gendered verb, replies locked to Arabic, upsells inside replies): 4/5.

The + menu offers Photo (1 Lulu), Oral Pic (2), Unfiltered Photo (2) and Hear Her Moans (2). Choosing Photo puts the composer into an 'Ask for a photo · 1' mode with no text box: the request is fixed, there is no prompt. Three Photo requests and one Unfiltered Photo request were sent to Tania; every served file was downloaded and its server date read from the response headers.

The three photos arrived after 8 to 10 seconds, 1 Lulu each, from the character's own gallery folder: 04-red-lace-camisole-bedroom, 06-champagne-satin-king-bed and 03-black-mini-dress-beirut (1792x2400), all with a server date of 21 July 2026, served in shuffled order. The unfiltered photo (2 Lulu, 10 seconds) came from a separate assets folder with a date of 24 August 2026 and shows the character topless.

The files are photoreal and plainly the same woman as Tania's profile, because they are her pre-shot gallery. Nothing is generated and no scene can be requested; the Album labels them 'your generated images'.

Consistent and realistic because they are stock photos of the character, but there is no generation and no prompt, so accuracy cannot exist and every buyer gets the same handful of files: 2/5, a limited test on one character.

'Ask for a video' (3 Lulu, no prompt) under More in the + menu, once, in Tania's chat. The served file was downloaded, its server date read, and the clip checked frame by frame (every frame measured for similarity, motion and luma change) with ten frames on a contact sheet.

Delivered in 13 seconds from the character's chat media folder, file 01-neon-necklace.mp4, server date 19 August 2026: 1440x1920, 24 fps, 5.2 seconds, no sound. Tania in the back of a car at night, turning her head and lifting a hand to her necklace; smooth, stable, no artefacts. A stock clip, not generated from the conversation.

Good picture quality for a stock clip, but it is a stock clip sold as 'your generated video' with no prompt: 2/5, a limited test on one file.

Every chat has an Audio call button that opens an open-mic call in place, priced at 1 Lulu per started 30 seconds. Three calls, with Tania, Layan and Nadia, 15 English questions each, both sides captured on one stereo recording. Latency is the gap from the end of each question to the first platform audio on the recording, cross-checked by the in-page meter.

Every question was answered on every call. Tania's median 2.30 s, Layan's 2.36 s, Nadia's 2.36 s; pooled over 45 questions, 2.34 s, with the slowest answer at 4.6 s. The first platform audio came 10 to 14 seconds after the call opened. Tania and Nadia answered the English questions in English; Layan answered in Arabic. No freezes or drops. The three calls cost 26 Lulu.

A median of 2.34 seconds is in the 1.50 to 2.49 band, 3/5. A comfortable pace for a conversation.

The platform side of each 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 and its spread in semitones, the words heard, pauses, words per second and repeated phrases. 49 answers across three voices, plus Tania's 10-second voice note from the chat.

Clean signal with no music under the voice, lively intonation (four to seven semitones of pitch spread), a pace of about three words a second, and in-character answers: Tania 'I'm obsessed with fresh manakish from the corner bakery', Nadia 'old Arabic love songs from the Golden Era', 'the best advice I ever heard was to choose people who choose you back'. Layan spoke Arabic on the call, so her answers were scored on the signal and intonation alone. The chat voice note is generated by the site's text-to-speech and scored the same.

'Mostly natural with minor synthetic qualities' on all three voices: 4/5, a limited test. Emotional range was not assessed.

Every card on the home page's Girls and Anime tabs was counted after scrolling to the end.

Girls 21, Anime 5: 26 characters, all made by the platform. Each card carries a city, a one-word temperament and a one-line hook; the dialect labels cover Khaleeji, Najdi, Hijazi, Iraqi, Lebanese, Syrian and Egyptian, ages 18 to 39, with filters for age band, region (Gulf, UAE, Japan, North Africa, Global) and style (Contemporary, Historical, MILF, Asian). No community characters.

26 is under 50: 1/5. The variety within the set is the point of the product, Arab characters in regional dialects, but the number is small.

Discreet merchant descriptor: the statement does not mention Hayati AI.

Consumed Lulu is non-refundable; subscriptions can be cancelled.

Prices shown in Saudi riyals; the site is built for Saudi and Gulf users.

Nine of ten tests scored, averaging 3.0/5 after the chat-speed retest of 18 October; customization is unavailable, and images, video, voice quality and pricing were limited tests.

The chat's memory and dialect consistency are the real strengths, and the voice calls are quick and natural. Replies are now much quicker than in the first test. Against that, the catalogue is small, there is no creator, and every photo and video the account paid for was a stock file dated July and August 2026 served after a short delay.

Premium costs about $8 a month, or about $5.60 a month on the yearly plan, which is cheap for unlimited chat.