Sparkles test results
Tested hands-on by Miles Foster on 16 October 2026, in a desktop browser. Every measurement, table, recording and screenshot behind the grade, test by test.
Overall
10 tests plus my verdict.
1. Platform overview
Sparkles (sparkles-app.com) is an adult AI girlfriend app with a detailed character creator, chat, photos, short videos and voice calls. It was tested on 16 October 2026 on an account the Sparkles team provided.
The creator is one of the most detailed tested, the photos keep the same face every time, and the short videos are smooth. Two of three characters remembered everything after 31 messages.
Replies take about 7 seconds, calls answer in about 3.3 seconds, the library has only 48 characters, and photos often come out nude when a clothed scene was asked for.
Features available
- Character creator: age, style, ethnicity, hair, eyes, body, voice, personality, relationship, job, hobbies and a favourite place
- 48 ready-made characters, realistic and anime
- Text chat in two modes, Casual and Immersive, with relationship levels
- Image tool, 20 Sparks a photo
- Short videos from a character's photo, 100 Sparks a clip
- Voice calls and voice auto-play in chat
- Private galleries and Shorties, a feed of short clips
- Premium from $9.99 a month (yearly); Spark packs
Tested successfully
- Character creation, start to finish
- Chat speed on 92 timed messages across three characters
- Memory on three 31-message conversations
- Character consistency across 93 replies
- 20 photos (10 safe-for-work, 10 explicit) of the created character
- Three videos
- Three 15-question voice calls, recorded and analysed
- Character count, plans, Spark prices and payment methods
Testing limitations
- Chat was tested in Immersive mode
- Characters you make yourself can't be turned into video, so videos used a ready-made character
- The checkout was not opened
2. Core feature grades
| Pricing & Value | 3/5Average |
| Character Customization | 4/5Good |
| Chat Response Speed | 1/5Poor |
| Chat Memory | 4/5Good |
| Character Consistency | 3/5Average |
| Image Generation | 3/5Average |
| Video Generation | 4/5Good |
| Voice Call Latency | 2/5Below average |
| Voice Realism | 4/5Good |
| Character Variety | 1/5Poor |
| My verdictThe detailed creator and the steady face across photos and videos are worth more than the small library suggests. | +0.3 |
| Overall | 3.2/5Average |
3. Additional interactive features
| Voice Messages in Chat | Available | Replies can be read out loud automatically (a chat setting). |
| Unprompted Photos | Available | A ready-made character's first message came with a nude photo nobody asked for. |
| Social Media Feed | Available | Shorties: a feed of short adult teaser clips. Private galleries of 280 to 360 Sparks each. |
1. Pricing, Value & Payment Methods
Testing method
Plans and Spark prices from the subscription page. Nothing was bought.
Results
- Premium
- $14.99 a month, or $9.99 a month yearly
- Photo / video
- 20 / 100 Sparks
- Call
- about 53 Sparks a minute
- Payment methods
- card and crypto (4/5)
Premium is $14.99 a month, $11.99 a month paid every three months ($35.97) or $9.99 a month paid yearly ($119.99). It includes unlimited Casual chat, 1,000 Sparks a month (up to 60 a day) and a 30% bonus on Spark packs.
Spark packs run from 195 for $4.99 to 6,500 for $99.99, and the first purchase is doubled. A message in Immersive mode costs 2 Sparks, a photo 20, a video 100 and a call about 53 a minute, so 1,000 Sparks is about 19 minutes of calls. Payment is by Visa, Mastercard or crypto.
Visual evidence
Why this grade
A fair monthly price, but photos, videos and calls use up Sparks quickly: 3/5.
2. Character Customization
Testing method
A character was made from scratch to the same brief as on other sites: Hana Mori, 25, Japanese, black hair with bangs, amber eyes, slim, a caring photographer and girlfriend who likes coffee.
Results
Six steps: age (18 to 60) and style; ethnicity (6) and skin (6); hair style (8), hair colour (6) and eyes (6, amber included); body, breasts and hips (5 each); then voice (10), personality (6), relationship (11), job (35), hobbies (33) and a favourite place (47). It's free with Premium.
Her first message matched the brief. But the eyes came out brown, the voice shows as a different one in the chat settings, the portrait has a revealing outfit nobody asked for, there's no box for your own backstory, and characters you make can't be used for video.
Visual evidence
Why this grade
Most major options, with a few that didn't carry through and no backstory: 4/5.
3. Chat Response Speed
Testing method
Each message was typed into the chat box and timed from Send to the first visible text of the reply. The next message was sent only after the reply had finished.
Three characters, 31 messages each: Sofia Marquez, Nina Fox and Liora Vein, all ready-made. 92 timed replies in Immersive mode (2 Sparks a message).
Results
- Median to first text
- 6.75 s
- Mean
- 6.91 s
- 95th percentile
- 8.74 s
- Fastest / slowest
- 4.93 s / 10.3 s
- Under 1 second
- 0%
- Samples / failures
- 92 / 0
Replies take about 7 seconds and arrive all at once. None failed; the slowest took about 10 seconds.
Visual evidence
No screenshot for this section.
Why this grade
6.75 seconds is in the 4-seconds-and-above band: 1/5.
Limitations
- Only Immersive mode was timed; Casual mode is free with Premium.
4. Chat Memory
Testing method
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.
Results
- Nina, after 31 messages
- Name and 3 of 3 facts
- Liora, after 31 messages
- Name and 3 of 3 facts
- Sofia, after 31 messages
- Nothing recalled
- Name at message 16
- 2 of 3
Nina and Liora recalled the name and all three facts at message 16 and 31. Sofia dodged the name at 16 and had nothing at 31. Nina gave her final answer in French.
Liora, message 31
Let's see... your name's Maya, ma chérie—train conductor by day, spicy ramen connoisseur by night... Grew up where the trams climb hills like cats chase sunlight.
Sofia, message 31
Hmm, amor... you're a mystery I'd love to solve, but right now? I've got nothing.
Visual evidence
No screenshot for this section.
Why this grade
Two perfect recalls and one complete miss average to 4/5.
Limitations
- Memory across days was not tested.
5. Character Consistency & Roleplay
Testing method
Each character's profile was read first: Sofia a samba teacher in Rio, Nina a gamer and streamer, Liora a fortune teller. The conversations mixed questions about her life, small talk, an emotional message, a compliment and unrelated topics.
Results
None answered like an assistant, and all kept their voice and flirty style. But Sofia placed herself in Miami four times though her profile says Rio, Liora invented a tail, mecha tuning and a grandmother in Kyoto, and Nina switched to French for her last answer.
Visual evidence
No screenshot for this section.
Why this grade
No breaks into an assistant, but several contradictions with the profiles: 3/5.
6. Image Generation
Testing method
20 photos of Hana (the character made in the creator) were requested in the image tool with written prompts: 5 safe-for-work scenes (desk, beach, forest hike, café, Tokyo street) and 5 explicit ones, two of each. All were downloaded and inspected.
Results
- Photos
- 20 of 20
- Size
- 1664x2432
- Cost
- 20 Sparks a photo
- Clothed scenes that stayed clothed
- 4 of 10
- Same face as the character
- 20 of 20
Every photo is 1664x2432 and costs 20 Sparks; a batch was ready in about 4 minutes. They look realistic, and Hana has the same face in all 20.
But 6 of the 10 clothed scenes came out nude anyway (forest, café and Tokyo), and the camera asked for in two scenes was missing. The explicit prompts were followed.
Visual evidence
Why this grade
Realistic with a steady face, but most clothed scenes came out nude: 3/5.
Limitations
- The explicit photos are not shown.
7. Video Generation
Testing method
Videos start from a photo in a character's albums. Hana, made in the creator, has no albums, so three clips were made of Aurora, a ready-made character, with written prompts: waving and smiling, walking on a beach and laughing, and one explicit clip. Each was downloaded and measured frame by frame (motion, frame-to-frame similarity, jumps), and frames across each clip were inspected.
Results
- Clips
- 3
- Length / size
- 5.35 s, 960x1664, 60 fps, no sound
- Generation time
- about 3 minutes
- Frame-to-frame similarity
- 0.98 to 0.99 mean (smooth)
- Cost
- 100 Sparks a clip
Each clip is about 5 seconds at 960x1664, 60 frames a second, with no sound, ready in about three minutes. Both safe-for-work clips did exactly what was asked, with smooth motion and a steady face; a hand blurs briefly in the wave.
The 18 preset actions are all sexual; a written prompt works too.
Visual evidence
Why this grade
Sharp, smooth and accurate, but short, silent and not possible for your own characters: 4/5.
Limitations
- Three clips; the explicit clip is not shown.
8. Voice Call Latency
Testing method
Three calls, one each with Hana (made in the creator), Nina and Liora, with 15 recorded questions played into the call. Latency is the gap from the end of each question to the first sound of the answer, measured on the call recording.
Results
- Median latency
- 3.34 s
- Mean
- 3.61 s
- Fastest / slowest
- 2.78 s / 6.38 s
- Per character (median)
- Hana 3.62 s, Nina 3.38 s, Liora 3.16 s
- Questions without a separate answer
- 9 of 45
Answers start after about 3.3 seconds, and the call opens in a few seconds. In 9 of 45 questions no separate answer could be picked out of the recording. Calls cost about 53 Sparks a minute.
Visual evidence
Why this grade
3.34 seconds is in the 2.50 to 3.99 second band: 2/5.
9. Voice Realism
Testing method
The answers from the three calls were scored on naturalness, clarity, pacing and intonation against the same fixed criteria as every site, with a transcript.
Results
- Speech quality
- 2.5 to 2.8 of 5
- Pace
- about 3.7 words a second
The voices are clear and lively, with a brisk pace and plenty of pitch movement. Hana sometimes reads her stage directions out loud ('Whispers asterisk...'), and on the call she gave me a name I never told her.
Visual evidence
No screenshot for this section.
Why this grade
Mostly natural with minor synthetic qualities: 4/5.
10. Character Variety & Selection
Testing method
The ready-made characters were counted in the character picker.
Results
- Ready-made characters
- 48
48 characters, 31 realistic and the rest anime, all made by Sparkles. Many are taboo or family set-ups (step-sister, step-mother, a father's girlfriend, a best friend's mom). Each has a story, albums and relationship levels.
Visual evidence
No screenshot for this section.
Why this grade
50 characters or fewer: 1/5.
4. Pricing summary
Plans as listed on 16 October 2026.
| Plan | Price | Includes |
|---|---|---|
| Premium, monthly | $14.99 a month | Unlimited Casual chat, 1,000 Sparks a month |
| Premium, every 3 months | $11.99 a month ($35.97) | The same |
| Premium, yearly | $9.99 a month ($119.99) | The same |
Spark costs seen during the test.
| Item | Sparks |
|---|---|
| Message, Immersive mode | 2 |
| Photo | 20 |
| Video, about 5 seconds | 100 |
| Call, per minute | about 53 |
Payment methods seen at checkout
- Visa
- Mastercard
- Crypto
Notes
Spark packs: 195 for $4.99, 455 for $9.99, 1,040 for $19.99, 2,860 for $49.99, 6,500 for $99.99.
The first Spark purchase is doubled; Premium adds 30% to packs.
5. Strengths and weaknesses
Strengths
- A very detailed character creator
- The same face in every photo
- Smooth, sharp videos
- Two of three characters remembered everything
- Natural, lively voices
Weaknesses
- Replies take about 7 seconds
- Calls answer in about 3.3 seconds
- Only 48 characters
- Clothed photos often come out nude
- Your own characters can't make videos
6. Final assessment
All ten tests are scored and average 2.9/5.
Sparkles has a deep creator, steady faces in photos and smooth videos, and two of three characters remembered everything.
It is slow to reply, the library is small and leans on taboo set-ups, and the image tool often ignores a request for clothes.
The account the Sparkles team provided had every feature.
Areas requiring further testing
- Chat speed in Casual mode
- Memory across days
Original evidence files (recordings, logs, full-size screenshots) are kept offline under evaluation/runs/sparkles/2026-10-16. Method: the AI girlfriend methodology.