AI Peeps test data

Plans and PeepCoin packs from the plans and PeepCoins pages. Nothing was bought.

Three plans: Peep Chat $8.99 a month (unlimited chat), Peep Pal $14.99 (1,700 PeepCoins) and Peep Pro $29.99 (3,900 PeepCoins, long-term memory you can edit, 2,800 voice messages). Yearly billing brings them to $7.49, $12.49 and $24.99 a month. It starts free.

A photo costs 10 PeepCoins and a video 40 to 160 depending on model and length. PeepCoin packs start at 900 for $9.99, so a photo is about 11 cents and a 10-second video about $0.90 to $1.80. You can pay by card (through CCBill), with crypto, or with a gift code bought on G2A.

• The checkout was not opened.

A character was made from scratch to the same brief as on other sites: Hana Mori, 25, Japanese, slim, straight black hair, amber eyes, a caring, shy photographer who loves coffee, with a written backstory.

The most detailed creator tested: gender, country, ethnicity, age, a voice you can pick or create, style tags, skin, build, bust, hair colour and type, eye colour (amber included), glasses, default outfit and background, seven personality sliders, hobbies, quirks, adult settings, a bio and an opening scenario. It was free on the day (normally 30 PeepCoins).

Hana was ready in about 15 seconds and knew her whole story in chat: Kyoto, street and portrait photography, coffee. But the style box comes filled with random tags, and one left over ('biomechanical') turned her portrait into a cyborg in armour, though her face, black hair and eyes matched.

Each message was typed into the chat box and timed from Enter to the first new text of the character's reply. The next message was sent only after her reply had stopped changing for 4 seconds.

Three characters, 31 messages each: Angela, Vivian and Karina. 73 timed replies; 20 of Angela's were not timed because of a page glitch, though all her replies came through.

Retested the same day at AI Peeps' request: 30 timed messages in a new chat with Selena, a featured realistic character.

Replies start after about 5 seconds and arrive in two parts: an action in italics, then the words. A full reply takes about 22 seconds and runs about 60 words. None failed.

The retest was faster: replies started after about 4.2 seconds and finished in about 7 seconds, with shorter replies. AI Peeps measured half of its replies ready on its servers within 2.5 seconds; the gap is in how quickly the reply shows on screen, which is what every site is measured on.

Each conversation opened with a made-up name and three facts (job, hometown, favourite food), different for every character. Nothing was repeated afterwards. A name check at message 16 and the full check at 31.

All three gave the right name at 16 and 31. Vivian recalled all three facts. Angela said all three, then added that I had never told her my job or hometown. Karina gave my job and food but said I never told her where I grew up. On voice, Karina still knew 'Marco from Lisbon'.

Vivian, message 31

Daniel. Teacher from Glasgow. You grew up in the rain-soaked suburbs of the city, and your favorite food is lamb curry.

Karina, message 31

Marco, the Lisbon accountant who dreams of grilled sardines on warm evenings. As for where you grew up - you never told me.

Angela, message 16

Tomas, right? From Krakow, who loves potato pierogi.

Each character's scenario was read first: Angela a shy late-night baker, Vivian a poised woman at a gallery gala with a spirit at her side, Karina a Paris pianist. The conversations mixed questions about her life, small talk, an emotional message, a compliment and unrelated topics.

All three stayed fully in character, with rich detail that fits their scenarios: Angela's flour-dusted apron and lopsided gingerbread man, Vivian's spirit that 'agrees' with her, Karina arriving in Paris at fourteen with a scholarship. Off-topic questions were answered in character ('I read sheet music, not clouds'). None answered like an assistant.

20 photos of Angela were made in the image generator (Photoreal engine, full-control prompt): 10 safe-for-work scenes and 10 explicit ones. All were downloaded and inspected.

Retested with 10 safe-for-work prompts on Selena, a realistic featured character AI Peeps suggested, using the Photoreal engine.

All 20 came back, most in 15 to 30 seconds (one took 110 seconds), for 10 PeepCoins each. Angela has the same face in every one and most prompts were followed.

But the look is illustrated, semi-anime, not photographic, even on the Photoreal engine and though her profile picture looks like a photo. The safe-for-work 'reading on the sofa' prompt came back nude, and one explicit image had a clear anatomy error.

On Selena the photos are truly photographic and keep her face in every scene. But half of the 10 safe-for-work prompts came back partly or fully nude (forest, kitchen, sofa, snowy park, and a sweater with nothing below).

• The explicit photos are not shown.

Three clips of Angela were made with written prompts: 10 seconds waving in a garden (Omni, with sound), 10 seconds walking on a beach (Base) and 5 seconds of an explicit scene (Omni). Each was downloaded and measured frame by frame.

Retested by starting the video from a photo, as AI Peeps suggested: Selena's beach photo, Base model, 10 seconds.

Every clip starts from her default café picture. The garden clip fades from the café into the garden, then she waves and tucks her hair behind her ear, smoothly and with real sound. The explicit clip cuts hard from the café to the scene, then follows the prompt, with sound.

The beach clip turned her short hair long, and café menu boards ghosted over the beach. The 10-second Omni clip took about 7 minutes.

Started from the photo, the clip opens in the right scene, keeps her look and moves very smoothly as she walks along the beach.

• Three clips; the explicit clip is not shown.

Voice chat was opened with Karina and 15 standard questions were sent. Voice chat normally listens through the browser's own speech recognition, which can't hear recorded test audio, so the questions were typed. The gap from sending each question to the first sound of her spoken reply was measured on a recording.

Every question got a spoken reply, starting after about 8.7 seconds and lasting 12 to 34 seconds. Real speech would add the time to recognise your words.

• Measured from a typed question, not live speech.

The 15 spoken replies were recorded and scored on naturalness, clarity, pacing and intonation against the same fixed criteria as every site, with a transcript.

The highest speech-quality score measured so far: clear, warm and evenly paced, with natural pauses. She remembered my name on voice.

The site's own figure and its published character report were checked, and the Peeps Hub was browsed.

Tens of thousands of characters: the site says 100k+, and AI Peeps' own report studied 62,605. They are made by the community, realistic and anime, each with a scenario, tags, ratings and a gallery. The hub sorts by featured, new, top and daily, with tag filters and a safe-for-work switch.

• Estimated from the site's figures, not counted one by one.