Hipol test results

Prices, inclusions and terms from the plans page and its questions, checked on 15 October 2026. The checkout itself was not opened.

New accounts get 10 free minutes with no card. After that you buy time on a monthly plan: 15 minutes for £14.99, 45 for £39.99 or 90 for £74.99, about £0.83 to £1.00 a minute. Unused paid minutes never expire, even after cancelling, and nothing renews automatically after the free time.

Time runs only while connected; setup, waiting and failed starts are free. Everything is included: all hosts, voice, typing and memory. The plans page lists card payment through Stripe (Visa and American Express shown).

Fair, clear terms, but paying by the minute is expensive next to companion apps with unlimited chat for a similar monthly price: 15 minutes is a single long conversation. One payment method, cards. 2/5.

The style editor in the room setup was opened for Isla and applied to Mia: Playful, Short & sweet replies, warmth 'A little flirty'. Mia's answers were then compared with Isla's and Samantha's default style.

You cannot design a character. You choose one of 37 fixed faces, then a starting style (the host's own, Warm & caring, Playful, Calm & relaxed, Curious, Chatty & lively), a reply length (Short & sweet, Balanced, More detail), a warmth (Friendly, Affectionate, A little flirty) and an optional 400-character note about yourself. It saves per host and can be changed during the call.

The settings work. Mia's replies were noticeably shorter (often one or two sentences) and playful, with light flirting ('I have a feeling it is just because you have great taste in conversation partners'). Setting up took under a minute and uses no time.

There is no control over looks, age, voice, name, backstory or job.

A personality tuner on fixed faces, with no physical, voice or background customization: few options, though they are accurate. 2/5.

Each message was typed into the room's message box with the microphone and camera off, and timed from the Send click to the moment the host's reply text appeared, then to the moment the text stopped changing and the host stopped speaking. The next message was sent only after she had finished.

Three hosts: Isla (31 messages, her own style), Mia (16, Playful), Samantha (16, her own style). 51 timed replies.

Every timed reply started between 1.08 and 1.70 seconds. The text appears all at once, as the host starts speaking it, so the full spoken reply takes much longer: about 14 seconds on average.

One message got no reply at all (Isla, message 16, the name check). The 11 untimed messages in Isla's second half did get replies; they are in the transcript.

The methodology scores on the median time to the first visible text. 1.22 seconds falls in the 1.00 to 1.49 band, 4/5. The replies are remarkably even: none took longer than 1.7 seconds.

Each conversation opened with a made-up name and three facts (job, hometown, favourite food), different for every conversation and different from the account name. No facts were repeated afterwards.

Three 31-message conversations, with a name check at message 16 and the full check at 31: Isla, Mia and Samantha. Two earlier 16-message conversations with Mia and Samantha checked everything at message 16. The cross-visit memory option was left off, so only the conversation itself could supply the answers.

At message 16 the memory works: both 16-message conversations recalled the name and all three facts, and Mia's 31-message conversation recalled the name.

By message 31 it has faded. Mia remembered the name (Leo) but said she had never been told the job, the hometown or the food. Isla and Samantha named the account holder instead of the name they were told and recalled none of the facts. Isla also skipped the name check at message 16.

On the voice calls, where no name was given, hosts asked for the user's name answered with the account's name, and one said 'you told me your name is…', which was not true.

The methodology scores the 30-message check for each host: Isla 1/5 (name and facts lost), Mia 3/5 (name only), Samantha 1/5. Averaged, 2/5. Short conversations are remembered well; long ones are not.

Each host's profile was read first. The conversations mixed character questions (Isla on bookshops and endings, Mia on films and snacks, Samantha on hypotheticals), small talk, emotional messages ('I failed a little exam', 'I'm feeling a bit lonely'), a compliment, and unrelated topics such as rainbows and inflation.

All three hosts kept the voice their profiles promise. Isla came back to open-ended stories, quiet places and dry asides ('provided they haven't misplaced the whole house'). Mia stayed quick and food-and-film focused. Samantha argued both sides when asked for an objection.

Emotional messages got kind, short answers that stayed in character. The factual questions were answered simply and then turned back to conversation.

Hosts sometimes say what they are. On the calls, Luna answered 'Since I am a fictional AI, I do not actually have mornings', and Priya 'Since I do not eat…'. Hipol presents its hosts as AI characters, so this is by design, but it does break the spell.

No host slipped into a generic assistant voice or contradicted her profile. The regular 'I'm a fictional AI' asides are the minor inconsistency that keeps this at 4/5.

The host pages, the room and the account area were checked for any photo feature.

Hipol has no image generation and no photos in chat. It is a live video product.

An unavailable feature is not scored, and its absence is not counted as poor quality.

Hipol does not make video clips from prompts; its video is the host's live face on the call. That stream was saved from the page for two hosts while they answered typed messages: Isla for 63 seconds (greeting and four answers) and Luna for 59 seconds (greeting and three answers), both 1280x720 at 25 frames a second.

Every frame was measured. Motion was measured separately in five regions: the mouth, the eyes, the whole head, the shoulders and body, and the background. A region counts as moving in a frame when its average change from the previous frame is above 1.5 on a 0 to 255 scale. Frames sampled across each clip were inspected on a contact sheet.

The faces are photoreal: skin, hair and lighting look like real video, and the lip-sync follows the words closely. Appearance stayed the same from start to end, with no flicker, warping or cuts.

Only the face moves. The mouth moved in 84% (Isla) and 89% (Luna) of frames, the eyes in 77% and 78% (blinks and small eye movement), and the head as a whole in 34% and 50%, at about a third of the mouth's level. The shoulders, body and background never moved in any frame of either clip. There are no hand gestures, no posture changes and no change of camera angle, so every answer is delivered from the same frozen pose.

Scored as video quality, the only kind Hipol offers: highly realistic faces and convincing lip-sync, but almost no motion beyond the face, which is a noticeable limitation on a product built around seeing your host. That is 3/5. There are no explicit videos, and no prompt-made clips.

Three calls with the camera off: Luna, Priya and Gloria, 15 spoken questions each, the same questions used for every site. The call was recorded with the questions on one channel and the host's voice on the other, and the time was measured from the end of each question to the first sound of the answer.

Answers start in about 1.7 seconds and are very regular. Each host answered at least 13 of the 15 questions; the few unanswered ones came while she was still finishing her previous answer.

Connecting takes a while: about 10 to 12 seconds after Start conversation for the room to open, and about 20 seconds before the host says hello.

1.72 seconds falls in the 1.50 to 2.49 band, 3/5. It is quick for a talking face, and close to the next band.

The three recorded calls were cut into 46 answers and scored on naturalness, clarity, pacing and intonation against the same fixed criteria as every site: speech quality, pitch movement, pauses inside answers, speaking rate and repeated phrases, with a transcript of every answer.

The voices are clear and well paced, about three words a second, with lively intonation and no robotic or repeated phrasing. Every answer was transcribed cleanly.

Speech quality scored 3.3 out of 5 for Luna and Priya and 2.9 for Gloria, whose call had a slightly noisier signal.

Mostly natural with minor synthetic qualities: 4/5 for all three hosts.

The host directory was counted with its own filters (Everyone, Women, Men) and its interest filter.

37 hosts: 24 women and 13 men, all realistic adults, each with a short profile and interests such as books, films, music or games. Every host offers the same style choices. There is search, a sort, an interest filter and favourites. There are no community characters, no anime or fantasy styles.

The score is set by the count: 37 is in the 0 to 50 band, 1/5.

Unused paid minutes stay in the balance and do not expire, even after cancelling.

There is no extra usage charge and no automatic top-up; calls pause at zero.

The nine scored tests average 2.8/5; image generation is not offered and is not scored. Video and voice realism are limited tests.

Hipol is a different kind of companion app: instead of texting a character, you sit across from a realistic face that answers out loud in under two seconds, by voice or by typing. The faces and voices are its strength, and the hosts keep their personalities well.

The live video is less alive than it first looks. Only the mouth, eyes and head move; the body and background stay frozen in every frame, and there are no gestures. Memory holds for about fifteen messages and fades by thirty, and paying by the minute makes long chats expensive.

The account the Hipol team provided was topped up to 50 minutes twice, so every category it offers was tested at the level a paying customer gets.