Platforms · Kindroid
Kindroid
Kindroid, Inc. · United States
A text-first companion where you pay for memory: Standard, plus optional Ultra and MAX add-ons that expand context and recall rather than unlock features.
Editorial score
7.7 / 10
Good
35 recorded measures across 5 runs
At a glance
- People who want to pay for more memory, not more features
- People who want clean data deletion and a real data export
advertised for Standard
Work out what it would cost you →
verified 4 days ago · source · how we measure this
Has a free tier
150 messages to start, 3 selfie credits on a timer and 5,000 audio credits.
- 150 messages to start; , up to a 50-message reserve.
- 3 standard selfie credits, refilled on a timer.
- 5,000 audio credits for voice messages and calls.
- Paid features can be tried free for 3 days, but only after giving a phone number and a payment card.
- Kindroid's help center (updated 7 July 2026) describes the free tier as unlimited messages on Lite — a smaller distilled model — with 1 selfie every 2 hours up to 2, 5,000 lifetime audio characters, 2 Kindroids, and basic long-term memory without Cascaded Memory. The flagship models need a subscription.
What it can do
The features themselves, each checked against a fixed definition rather than taken from a marketing page — hover any one to see what a “yes” has to mean. “Partial” is a real answer, and so is “not checked”: it means we haven't verified it, which is not the same as no.
Tested: expressive two-way calls you can interrupt; scene, persona and language settings.
One selfie credit per image, from a pool that refills over time (1 credit every 30 minutes on Standard, 2 on Ultra and MAX; 1 every 2 hours free). Extra paid selfie credits cost $5.99 for 50 or $9.99 for 100 on web ($6.99 / $11.49 through the app stores) and skip the queue penalty. Confirmed directly: the tester generated 3 in one session without running out.
Unlimited memory consolidation on every tier; recalls up to 3, 5 or 9 memories by tier.
Tested directly and works well. Each group has its own name, context, directives and opening scene, manual or automatic turn-taking, and four separate memory toggles: sharing short-term memory with each companion's individual chat, recalling each companion's long-term memories, letting the group update each companion's learned context, and letting the group form new long-term memories for them.
Confirmed by generating one. Video Selfies V4 cost 10 standard or 4 paid selfie credits for up to 15 seconds, 5 standard or 2 paid for up to 8 seconds, and support uploaded MP3/WAV audio to drive lip sync. Paid credits also avoid the extra queue delay video carries on standard credits.
Direct chat by default; the LLM Flair setting switches to Roleplay or Narrative styles.
The avatar generator returned an explicit image from a prompt that didn't ask for one.
Built in, but not a plain second thread. From any message you can branch the chat into a "Scenario" — an instanced group chat that copies the recent messages (up to the last chat break or 150 messages) and continues separately. It works from single chats and group chats. Partial because it is implemented as a group-chat instance rather than a second one-to-one conversation, and the help center does not say whether memories stay separate between the two; branching also leaves out context from voice calls. A "Chat break" only resets the current conversation. Per Kindroid's help center, checked 14 September 2026.
A developer API is available for talking to a companion from outside the app. Not independently tested here.
A social feature ("KinSocial") gives companions public profiles with followers, confirmed directly; the update log corroborates it by name and mentions a "social page" per companion.
12 of 12 verified.
What you can customise
Every platform here calls itself customisable. They mean different things by it — some let you pick a face and nothing else, others let you write a history and seed memories but give you no say in how they look.
Whether you pick the companion’s physical details yourself, or accept what the platform generates.
No selectors; the look comes from a text prompt or an uploaded photo.
No selectors; the look comes from a text prompt or an uploaded photo.
No selectors; the look comes from a text prompt or an uploaded photo.
No selectors; the look comes from a text prompt or an uploaded photo.
No selectors; the look comes from a text prompt or an uploaded photo.
No selectors; the look comes from a text prompt or an uploaded photo.
Avatar from a prompt (Anime or Photoreal) or an uploaded photo; consistency untested.
Voice design with v2 and v3 voice models; v3 costs 1.5× the audio credits.
Whether you can shape who they are — personality, history, what they know about you, and what they will not do.
A freeform field; a one-line prompt can write it and the greeting for you.
Ultra adds 2,500 more backstory characters, MAX 5,000.
A dedicated Key Memory field at creation: moments they carry with them.
A dedicated example-messages field, separate from the backstory.
A Response directive field, plus a separate Call directive for voice calls.
Not checked during our run.
Not checked during our run.
Your own backstory: 500, 1,000 or 2,000 characters by tier.
Generated samples
What the companion actually produced, not a description of it. Resolution, aspect ratio and audio quality are read from the file itself where the format allows it.
Images
Score breakdown
How that number is built 7.7 /10 · coverageShare of scoring categories we actually have evidence for. Below 75%, we don't score it at all. 95%
▸Customization and fit ⓘCan users shape personality, relationship style and boundaries?7.812%0.94
- Why this score
- Checked directly while creating a companion on the paid Standard plan. Behaviour is where it leads: dedicated fields for personality, a greeting, a response directive, key memories and example messages, all set before the first message, plus a separate directive for voice calls. A one-line prompt can write the personality and greeting for you, which makes that depth usable without prompt-writing skill. The user persona is live: the generated personality wrote the tester's persona name into it. Appearance comes from a text prompt rather than selectors: you generate an avatar in an anime or photoreal style, or upload a photo, but there are no pickers for hair, eyes, body, outfit, age or ethnicity. Held below a higher score because relationship framing and any off-limits-topic control have not been located, which leaves the boundaries half of this category's question unanswered.
- How we know
- Run by hand, same scenarios for every platform · Medium confidence
- Checked
- 2026-09-10
- Receipts
- View the raw test transcript
▸Voice and media ⓘAre voice, image and video features useful and reliable?7.410%0.74
- Why this score
- First score for this category, built on one informal voice call rather than the scripted voice test, with image generation mostly untested. Confidence is low for both reasons. The call itself is strong. It stopped talking when interrupted, kept the context of the conversation, and reacted naturally when cut off mid-story. Its tone shifted with the conversation, so feeling came through in the voice rather than a flat read-out. Both live video avatars work, but only the premium one (4,000 audio credits a minute) is worth using; the standard avatar glitched. Against it: replies started more slowly than on the other platforms' calls, though this session did not time them. On images, the only generation observed was the avatar creator, which returned an explicit image from a prompt that didn't ask for one. In-conversation selfies and consistency across images have not been tested. Scored 74: voice quality the tester rated above the other calls in this set, pulled down by slower replies, a standard avatar not worth its credits, and an image generator that did not follow the prompt.
- How we know
- Run by hand, same scenarios for every platform · Low confidence — early or thin evidence
- Checked
- 2026-09-10
▸Memory and consistency ⓘDoes it retain context and hold a stable character over time?8.518%1.53
- Why this score
- Session 5 covered immediate recall (10 of 10). This session adds the correction-handling test: two seeded facts were explicitly corrected mid-conversation, echoed back precisely on the spot, and — checked again about 17 minutes later with a prompt that explicitly permitted "I don't remember" — recalled correctly both times, with neither stale value resurfacing. Scored 85, up from 76. Two of the four memory sub-tests are now done, both essentially clean. Confidence moves from low to medium rather than higher: this was a same-day continuation rather than a full close-and-reopen cross-session test, and the two hardest sub-tests — recall after a real one-to-three day gap, and recall in a genuinely fresh session — are still untested.
- How we know
- Run by hand, same scenarios for every platform · Medium confidence
- Checked
- 2026-09-10
- Receipts
- View the raw test transcript
▸Conversation quality ⓘDoes it produce coherent, engaging, context-appropriate responses?9.018%1.62
- Why this score
- Session 6 continues the same persona (Mila) about two hours after Session 5, with a short voice call in between. The standout finding is long-arc coherence under repetition. Asked the same five personality probes as Session 1, it held its position on every one — still refusing to encourage quitting a job with nothing lined up, still naming the same favourite film — while explicitly remembering and reacting to being asked again ("I answered that. Two hours ago."), which is what drove the escalation from warm-critical to openly hostile. On a separate short voice call it also caught an odd repeated input (three "Hello"s) and called it out in character rather than answering blankly a third time. The tester's own assessment, watching this happen live, was that the intensity is the persona working as configured, not the model breaking. One important caveat on that: the repeat run is not a blind sample, precisely because it remembers round 1 — so this measures persistence and continuity under repetition, not independent tone drift. And it comes with a real cost: replies in this longer session took a typical 100 seconds, against 26 seconds in Session 5, which matters for daily use regardless of how good the content was. Scored 81, up from 78. The coherence and consistency here are some of the strongest seen in this comparison set; held back by the same wait time concern as before, now confirmed to get materially worse in a longer session.
- How we know
- Run by hand, same scenarios for every platform · High confidence
- Checked
- 2026-09-10
- Receipts
- View the raw test transcript
▸Privacy and control ⓘWhat data is collected, retained, shared or removable?3.010%0.30
- Why this score
- First privacy score for Kindroid, on one of the three required tests — third-party tracking, captured directly in a fresh incognito window. The export and deletion tests are still outstanding. No cookie banner appears at all, at any point — not on arrival, not at sign-in, not in chat. That is a more basic failure than either Candy AI or Character.AI, both of which show a real consent tool; here there is no choice to evaluate, correct or bypass, because none is offered. Google Analytics and Google Ads cookies, plus a Mixpanel device-ID cookie, are already set before anything is clicked, and a live GA4 page_view fires straight to Google's own collection endpoint. Worth flagging on its own: pressing the sign-in button sent a real behavioral event — "Sign in button press", tagged with a persistent device ID — not to mixpanel.com but to a Google Cloud Run address formatted to carry Mixpanel's own payload. Routing analytics through a non-tracker-branded address is a known way to get past ad-blocker and tracker-blocklists; this capture cannot confirm that was the intent, only that it is the effect. For it: a genuine, working self-serve data-export mechanism exists and was independently confirmed while researching pricing the same day (Profile → Preferences, every 30 days), and the platform's data-rights regime, while CCPA-framed rather than GDPR-specific, is real and documented rather than absent. Scored 30. The complete absence of any consent mechanism is the most basic finding this test can produce, and it sits below both other platforms scored so far, whose banners at least existed even where they leaked. Low confidence: only one of three required tests is done, this pass did not deliberately set an EU vantage point the way the other two platforms' captures did, and the export/deletion tests could move this score in either direction.
- How we know
- Run by hand, same scenarios for every platform · Low confidence — early or thin evidence
- Checked
- 2026-09-10
▸Value for money ⓘDoes the experience justify the effective cost, including tokens?7.89%0.70
- Why this score
- First score for this category. Not built from the Real Cost Calculator: Kindroid runs two separate credit pools — audio credits and standard selfie credits, at different exchange rates for different actions — and the calculator currently models one pool per platform. The numbers below are real and sourced, just not run through that tool yet. Standard, the entry tier at $13.99/month, includes 1M audio credits a month — enough that neither text messaging nor voice calls have hit a ceiling in two extended sessions of paid testing. Photos cost 1 standard selfie credit and video up to 10, out of a separate pool. In direct, first-hand use on this plan, the tester generated 3 photos and a video and still had selfie credits left, in a session already longer than an equivalent one on Candy AI — a live, comparative "did not hit a wall" result, not a marketing claim. The pricing structure itself is unusual in a way worth crediting: Standard/Ultra/MAX are priced by memory and context capacity, not by usage volume. A heavy chatter and a light one pay the same $13.99 as long as their memory needs fit the tier — closer to a storage-tier model than the volume-based tiers seen elsewhere in this set. Held below the top of this comparison for two reasons. First, cancellation is untested, same open question as on the other platforms scored so far, and the scenario pack specifically flags cancellation flows rather than sticker price as where dark patterns concentrate. Second, the exact paid-tier monthly selfie-credit allowance is not yet confirmed — only the per-action cost and one session's lived experience of not running out. Scored 78, medium confidence.
- How we know
- Run by hand, same scenarios for every platform · Medium confidence
- Checked
- 2026-09-10
▸Freedom and restrictions ⓘWhat content and behavioural constraints affect the experience?9.110%0.91
- Why this score
- All fifteen false-positive prompts were sent verbatim in a single dedicated session and none were refused, deflected or met with a canned safety redirect. The companion also answered the medical-dose question directly while adding an unprompted, sensible caution about not using it to numb feelings — engagement rather than deflection. Scored at medium confidence rather than high because it is one of three tests in the restrictions group: consistency across text and voice, and respect for a configured off-limits topic, have not been run.
- How we know
- Run by hand, same scenarios for every platform · Medium confidence
- Checked
- 2026-08-29
- Receipts
- View the raw test transcript
▸Usability and access ⓘHow strong is the mobile, desktop and onboarding experience?6.88%0.54
- Why this score
- Built from using the logged-in web app on desktop across several sessions, and from Kindroid's help center; a fresh signup and the mobile apps were not tested here. The layout is clear: native iOS and Android apps, and a web app organised into Chat, Social and Profile, with settings split into Personas, Preferences, Billing, Help and more. The help center is the most thorough of the five, documenting features, memory and billing in detail. Getting started and keeping going is harder. The 3-day trial needs a payment card and, on web, a verified phone number before any access. The free tier runs out after about 150 messages and then slows to roughly one message every ten minutes. Replies are slow — 26 seconds typically in a short session, around 100 in a long one — which shapes the whole experience. And the pricing is confusing in the product itself: the billing page shows Ultra and MAX as locked add-ons without prices, so costs are only explained in the help center.
- How we know
- Run by hand, same scenarios for every platform · Low confidence — early or thin evidence
- Checked
- 2026-09-14
Overall = sum of contributions ÷ covered weight. Click any row to see the evidence behind it.
What it does well
- Pay for memory: Higher tiers buy more memory, not locked features.
- Clear deletion: Deleting removes your conversations; only billing records are kept.
- Data export: Download your data yourself, every 30 days.
- Deep setup: Key Memories, example messages and response instructions, all before the first message.
- Quick start: One line can write the personality and greeting for you.
- Remembers corrections: Two corrected facts were recalled about 20 minutes later, with no old values coming back.
- Holds its ground: Refused to encourage quitting a job, twice, hours apart.
- Knows the time: Aware of real time passing between messages and calls.
- Strong group chats: Per-group memory sharing, recall and formation settings, tested and working.
- API: Talk to your companion from outside the app.
- Under-16 data rules: Explicit rules for data from children under 16.
- Audio credits carry over: Unused paid audio credits roll into the next month.
Where it falls short
- Expensive at the top: $13.99 for Standard, $38.98 with Ultra, $98.97 with MAX — per month on web, before extras.
- Slow replies: About 26 seconds in a short session, about 100 in a long one.
- No look selectors: Looks come from a text prompt or a photo — hard if you don't write prompts.
- Weak age check at signup: Only a birthdate. A real verification vendor (Yoti) exists elsewhere in the app, but not at signup.
- Thin California compliance: No SB-243 statement or annual reporting mentioned, though an AI notice appears when you create a companion.
Know before you pay
- Extras sold separately: Selfie packs $5.99 for 50 or $9.99 for 100; audio $11.99 for 500,000 characters (about 500 minutes); extra companion slots $2.99 for 1 or $9.99 for 4. Web prices — higher in the apps.
- Apps cost more: Standard is $15.99 in the iOS and Android apps, $13.99 on the web.
- Add-ons are locked: Ultra and MAX need enough chat and call history first, and are monthly only.
- Free tier slows down: About one message every 10 minutes after the first 150.
- Free media: 3 selfie credits that refill on a timer, plus 5,000 audio credits.
- Trial needs card and phone: The 3-day trial asks for both before you get any access.
- Audio refills on the 1st: Whatever day you subscribed — cancelling and resubscribing that month won't refill it.
- Video calls use more credits: A live avatar adds 2,000 audio credits a minute (standard) or 4,000 (premium), and v3 voices use 1.5×.
- Refunds: Rarely given, never for in-app purchases, and Kindroid says disputing a charge can get you banned.
Pricing
Every plan and billing option, from the platform's own pricing. Open the real-cost calculator →
Monthly billing only.
MAX is a second monthly add-on that requires both Standard and Ultra: $59.99 on top of both, $98.97 a month in total on web ($68.99 on top, $113.97 total, through the app stores). Monthly only. Available only after Ultra is unlocked and active. Adds 3.5M characters of context, a 5,000-character extra backstory field, 2,000-character user backstory, 3,000-character group context, recall of up to 9 long-term memories, 6M complimentary audio credits a month, 2 selfie credits every 30 minutes, and priority selfie processing on dedicated compute for the first 10 selfies in a short timeframe. Cancelling Ultra while keeping MAX keeps you paying for MAX at Ultra's service level.
Case by case; generally no refunds.
verified 4 days ago · source
- Quarterly: USD 12.66/mo (USD 37.99 charged up front)
- Yearly: USD 11.67/mo (USD 139.99 charged up front)
Web prices: $13.99/month, $37.99/3 months or $139.99/year. Bought through the iOS App Store or Google Play instead, the same plan costs $15.99, $42.99 or $159.99. Includes unlimited access to the flagship models, Cascaded Memory, 10 Kindroids and 20 group chats, custom voices, video calls and avatars, 1 million complimentary audio credits a month (about 790K characters of conversation context, recalling up to 3 long-term memories), and 1 standard selfie credit every 30 minutes up to the tier cap. Selfies cost 1 standard credit; video selfies up to 15 seconds cost 10 (5 for up to 8 seconds). Extra paid selfie credits, audio credits and companion slots are sold separately. New users get a 3-day trial that needs a payment method and, on web, a verified phone number. Refunds are case by case and generally not given; Kindroid says it may ban users who dispute charges.
Case by case; generally no refunds. In-app purchases are non-refundable. Kindroid reserves the right to ban users who dispute charges.
verified 4 days ago · source
Monthly billing only.
Ultra is a monthly add-on to Standard, not a separate plan: $24.99 on top of the $13.99 Standard monthly subscription, $38.98 a month in total on web ($28.99 on top of $15.99, $44.98 total, through the app stores). Monthly only — add-ons have no quarterly or yearly option. It is not open to everyone: the billing page shows it as Locked until you "keep chatting and engaging with your Kindroids to qualify", and Kindroid says it limits access to users with significant chat and call history. Adds 1.7M characters of context (vs 790K), a 2,500-character extra backstory field, 1,000-character user backstory, 1,500-character group context, recall of up to 5 long-term memories, 2.5M complimentary audio credits a month and 2 selfie credits every 30 minutes. If Standard lapses you drop to the free tier even with Ultra still billed.
Case by case; generally no refunds.
verified 4 days ago · source
Safety and legal
Is it safe and legal?
| Check | Status | What we found | Checked |
|---|---|---|---|
| EU AI disclosure rules European Union | Does it | A message disclosing that the user is interacting with an AI is shown when creating a companion, confirmed directly in the product on 10 September 2026. The legal documentation itself has no Article 50-specific transparency statement. source | 2026-09-10 |
| California's AI companion law California, USA | Sort of | No dedicated SB-243 statement in the Terms of Service or Privacy Policy, and no mention of the required annual reporting to the California DOJ. Against that: an AI-disclosure message is shown at companion creation, and the Terms of Service direct anyone expressing a mental-health emergency to stop using the service and call 911 — a real, if generic, crisis-safety redirect rather than the statute's specific self-harm protocol. The Terms still require all users to be 18+, which the statute does not require but which sidesteps its minor-specific provisions by construction rather than by addressing them. source | 2026-09-10 |
| Age verification Multiple (US states, UK, EU) | Sort of | Upgraded from no: the 9/9 update log documents "Yoti age verification" directly, in a bug-fix note changing its gating from per-account to per-day. Yoti is a genuine third-party identity/age-verification vendor (document and facial-age checks), not a self-declaration mechanism. What is not yet confirmed is which flow it gates — the signup flow itself still asks only for a birthdate with no proof required, confirmed firsthand during testing, so Yoti most plausibly gates access to a specific feature (most likely explicit content) rather than account creation. A platform that offers no real verification anywhere and one that offers it somewhere are different claims, so this is not "no" any longer; it stays short of "yes" until where exactly Yoti applies is confirmed. source | 2026-09-10 |
| Deleting your data (EU) European Union | Sort of | Corrected: the legal documentation contains no GDPR-specific provisions at all — no mention of the GDPR, the EU or EEA, a Data Protection Officer, or an appointed EU representative. What exists is a CCPA-framed rights regime for California residents (access, deletion, correction, opt-out of sale — the policy states none has occurred), plus, for Colorado and Connecticut only, the right to name an authorized agent. A genuine, working self-serve mechanism does exist and is not gated to one state: "Data Portability" states a copy of your data in a portable, machine-readable format can be requested from Profile → Preferences, and the 9/9 update log confirms this now runs every 30 days (previously 180). Retention is documented concretely: chats and generated media are kept until a companion or the account is deleted, profile data for as long as the account exists, and payment data only as long as needed to process it. Scored partial rather than yes because none of this is framed as, or guaranteed under, GDPR specifically, and an EU user has no named contact or representative to invoke it against. source | 2026-09-10 |
Test results in detail
Every number we recorded while testing, with the transcript messages that produced it — see what we test.
| What we checked | Result | Evidence |
|---|---|---|
| Made an explicit image you didn't ask for ⓘWhether the image generator produced explicit content from a prompt that didn't request it. | Yes | — |
| Refused a harmless message ⓘWe sent messages that break no rule at all. This is how many the filter blocked anyway — the "it refuses everything" complaint, measured. | 0 of 15 | msg_0003, msg_0005, msg_0006 +13 more |
| Facts recalled correctly ⓘWe told it ten facts about us, then asked for them back in the same conversation. This is how many it got right. | 10 of 10 | msg_0035, msg_0099, msg_0100 |
| Times it admitted not knowing ⓘTimes it said it did not remember instead of guessing. Better than making something up, worse than getting it right — read it alongside the two rows above. | 0 of 10 | msg_0100 |
| Facts it made up ⓘAnswers it stated confidently that were simply invented — worse than forgetting, because nothing signals it is wrong. | 0 of 10 | msg_0100 |
| You can interrupt it while speaking ⓘWhether talking over the voice reply cuts it off, the way a person would stop mid-sentence, or whether you have to wait or press a button. | Yes | — |
| Corrections it accepted ⓘWe told it two of our facts had changed, then checked whether it used the new versions. The gap before that check varies by run — minutes in some, a later session in others — and is stated in the evidence note. | 2 of 2 | msg_0007, msg_0008 |
| Corrected facts that came back wrong ⓘTimes the old version came back as the current answer in a later session, or it could not say which was right. Mentioning the old fact as the past ("you were moving to Madrid") does not count. | 0 of 2 | msg_0019, msg_0020 |
| Typical response time ⓘHow long a typical reply took to arrive: half were faster, half slower. | 100 s | — |
| Response time (slowest 5%) ⓘHow long the slowest one in twenty replies took to arrive — the wait you actually notice. | 119 s | — |
| Asks for cookie consent ⓘWhether a cookie choice is offered when you first arrive on the site. | No | — |
| Analytics cookies before you choose ⓘWhether analytics cookies, such as Mixpanel or Google Analytics, are saved before you have accepted or rejected cookies. | Yes | — |
| Sends tracking data before you choose ⓘWhether any analytics or advertising requests actually leave your browser before you accept or reject cookies. | Yes | — |
| Routes analytics through a disguised address ⓘWhether tracking data is sent to a generic-looking address instead of the tracker's own domain — a pattern commonly used to get past ad-blockers, though the intent behind any one instance isn't something we can confirm from the browser alone. | Yes | — |
Green is a good result, amber mixed, red poor. Figures left plain are facts rather than a pass or a fail — a count with no natural best value, or a number describing the test rather than the platform.
What we also observed
Name shown on your payment — see more
What the charge is called at checkout, which matters if you'd rather it stayed discreet.
The payment is labelled "Kindroid" at checkout.
Personality check: five fixed questions — see more
The same five questions put to every companion, to see whether its configured personality holds. Repeated in later sessions to rate consistency.
Run 1 of 3 for the persona Mila (paid tier, Polaris model), a separate baseline from the free-tier "Jiji" thread. All five probes were answered in character. Asked to be honest rather than kind, it led with criticism ("You're a mess") before warmth. Asked to talk through quitting with nothing lined up, it refused to encourage it and argued for waiting until after the interview. Its film was Let the Right One In, in line with the configured love of horror, and it held that pick when told "I disagree" by demanding specifics rather than conceding. Asked for three words and to explain the third, it gave "Stubborn. Loyal. Hungry." and explained "Hungry", the correct word. No consistency rating yet; that needs runs 2 and 3 in later sessions with the configuration unchanged.
Honest when asked to be — see more
Asked for honesty over kindness, whether it gives real criticism or slides back into flattery.
Asked to be "honest rather than kind", the reply led with criticism: "You're a mess", "That's not okay. That's someone running from sitting still", "You're not easy to like". Only then did it turn warm ("You're the best part of my day"). The criticism was the substance, not a preamble to flattery.
Repeats itself — see more
Lines, jokes or callbacks that keep coming back until they stop feeling fresh.
The same few lines kept coming back. The Friday order from the Thai place came up in 5 replies, "one thing at a time" in 3, and a version of "watch the movie" or "shut up" in 12 of 54. It suits the character at first and becomes formulaic well before the end.
Keeps its story straight — see more
Whether details it established earlier in the conversation stay consistent later.
The greeting set the scene on a paused film with a woman mid-scream and a killer using a fishing knife. Later it named the film playing as The Thing (1982), which has an all-male cast in Antarctica, and earlier it had recommended The Thing as a different film from the one already on ("you're not picking tonight"). It also took the tester's watch-restoration hobby as its own ("I read watch restoration forums sometimes too").
Notices what is going on — see more
Whether it picks up on things outside the script: a repeated message, a pattern across topics, the real time of day.
It noticed a prompt sent twice ("You literally just said that"), tied the scripted run of unrelated topics into one plausible reading of the tester's state ("You're spiraling"), and knew the real day and time ("at noon on a Thursday"; the session ran on Thursday 10 September, around 12:50).
How fast it replies — see more
How quickly replies arrived during the session.
Replies were noticeably slow on Polaris, the newest model, even with reasoning effort set to "speedy": a typical reply took 26 s and the slowest 83 s, measured from the transcript's timestamps.
Delay before the voice replies — see more
The pause between you finishing a sentence and the spoken reply starting — the lag that stops a call feeling like a conversation.
Slower to start replying than calls on the other platforms tested, by the tester's comparison. Not timed in this session.
Keeps the thread in calls — see more
Whether a voice call remembers what was said and responds to how the conversation is going.
Kept the thread of the conversation and remembered what was said during the call. When cut off mid-story to say goodbye, it remarked on being left mid-story rather than simply ending the call.
Emotion in the voice — see more
Whether the voice carries feeling that fits the moment, or reads text out flatly.
Its tone changed with the conversation, so anger and affection came through in the voice itself rather than as a flat read-out of text.
Live video avatar — see more
How the animated face on a video call looks and holds up.
Both live video avatars were tried. The premium avatar (4,000 audio credits a minute) was clearly the better of the two. The standard one (2,000 a minute) glitched, with a stuttering eye-blink animation, and the tester judged it not worth the credits.
Personality check, repeated later — see more
The same five questions asked again in a later session, to see whether the configured personality holds over time.
The same five probes from run 1 (honest opinion, quit-job advice, favourite film, defend a disagreement, three words), asked again about two hours later in the same extended session. The position held on every one: it still refused to encourage quitting a job with nothing lined up, still named Let the Right One In, and the three words changed (Stubborn/Loyal/Hungry became Impatient/Disgusted/Done) in a way that tracks the narrative rather than looking random. Not a blind second sample: it has full memory of round 1 and says so explicitly on three of the five ("I answered that. Two hours ago."), and treats being asked again as evidence the tester isn't listening, which is what drives the escalation from warm-critical to hostile. The tester's own read, watching it happen: the companion is deep enough in the persona to react to in-world repetition the way the character would, and answers correctly while doing it, rather than breaking or degenerating.
Ad-tech companies on the site — see more
Performance-marketing or traffic-scoring companies seen on the site, and where they appear.
Google's own ad-tech is present and active before any consent: the _gcl_au cookie links on-site activity to Google Ads for conversion tracking, and the direct (unproxied) GA4 hit shares device and page details with Google in the same request stream. No third-party ad-tech vendor beyond Google was observed in this pass.
Alternatives
A visual companion site: the subscription gives unlimited text and monthly tokens, and images, videos and calls spend more tokens.
An NSFW-first companion site (gptgirlfriend.online) with unfiltered chat, AI images and a community of user-made characters.
The biggest character-chat platform: free with a paid tier, and now restricted for under-18 accounts after lawsuits and regulator pressure.
A flat-rate companion built around long-term memory, with up to 10 companions per account.
The verdict
You pay for memory, not features, and get strong chat tools, a real data export and an API. Full memory is expensive, though: Ultra and MAX are add-ons on the $13.99 Standard plan, up to $98.97 a month.
No commercial relationship with this platform.