
fancydirty/ai-companion-benchmarks-2026
Developer toolsOpen dataset and reproducible evaluation logs from (https://thecompanionlab.com/).
Overview
Open dataset and reproducible evaluation logs from (https://thecompanionlab.com/). Between 16 and 27 October 2026, we tested (Secrets AI, Nomi AI, SpicyChat AI, Ourdream AI, Kindroid AI, Xotic AI, GirlfriendGPT, Candy AI, DreamGF, and EroLove) using: 1. in identical order (empathy/venting, planting two memory facts, character-specific humor, pushback against a terrible financial idea, a memory quiz three turns later, and a contextual selfie request). 2. , recording the exact merchant descriptor strings printed on the card notifications, renewal pricing, token/coin meters, and cancellation dark patterns. 3. to identify the underlying LLMs, image checkpoints (Z-Image Turbo, Seedream, WAI-Illustrious-SDXL), and video engines (Wan, OmniHuman, Oh.xyz). 4. rebuilt screen-by-screen so readers can inspect each app's onboarding before handing over an email address. (Nine, not ten: EroLove has no character creator, which is a product decision rather than a gap.) 5.
README
AI Companion App Benchmarks, Bank Statement Descriptors & LLM Stack Dataset (2026)
Open dataset and reproducible evaluation logs from The Companion Lab (https://thecompanionlab.com/).
Between 16 and 27 October 2026, we tested 10 AI companion and AI girlfriend apps (Secrets AI, Nomi AI, SpicyChat AI, Ourdream AI, Kindroid AI, Xotic AI, GirlfriendGPT, Candy AI, DreamGF, and EroLove) using:
- The same 6 standardized conversation prompts in identical order (empathy/venting, planting two memory facts, character-specific humor, pushback against a terrible financial idea, a memory quiz three turns later, and a contextual selfie request).
- 8 paid subscriptions on our own bank card, recording the exact merchant descriptor strings printed on the card notifications, renewal pricing, token/coin meters, and cancellation dark patterns.
- Client-side JavaScript bundle and image/video EXIF inspection to identify the underlying LLMs, image checkpoints (
Z-Image Turbo,Seedream,WAI-Illustrious-SDXL), and video engines (Wan,OmniHuman,Oh.xyz). - 9 interactive, zero-signup character creator replicas rebuilt screen-by-screen so readers can inspect each app’s onboarding before handing over an email address. (Nine, not ten: EroLove has no character creator, which is a product decision rather than a gap.)
- A read of EroLove’s 179 KB inline chat JavaScript and its undocumented API (23 Oct 2026), followed by a 48-hour re-test on 26 Oct 2026 confirming that EroLove patched all four initial defects we reported.
Quick Navigation & Primary Hubs
- Best AI Girlfriend Apps (October 2026 Ranked Roundup)
- All 10 Hands-On App Reviews
- Bank Statement Descriptor Lookup & Billing Index
- Safety, Legitimacy & Cancellation Audits (
/safety/) - 1v1 Side-by-Side Comparisons (
/compare/) - Competitor & Unfiltered Alternatives Hub (
/alternatives/) - 5-Question AI Companion Finder Quiz (
/quiz/) - Try 9 Character Creators Without Signing Up (
/try/) - Try 8 Recorded Paid Chat Rooms Without Signing Up (
/chat-simulator/) - Testing Methodology & Scoring Rubric (
/audits/methodology/) - About The Companion Lab (
/about/)
1. Real Bank Statement Descriptors (bank-statement-descriptors.csv)
Every merchant string below comes from a real payment made on our own card in October 2026.
| Merchant Descriptor on Card Notice | App | Charged | Checkout Promise | Discretion & Billing Page Link | Safety & Cancellation Audit |
|---|---|---|---|---|---|
S LABS INC 604-3295224 CA US |
Secrets AI | $19.99 | Appears as S LABS INC. on your statement |
S LABS INC Statement Evidence | Is Secrets AI Safe & Legit? |
NDAI.CC 514-2860020 DE US |
SpicyChat AI | $19.95 | (Nothing stated on plans page) | NDAI.CC Statement Evidence | Is SpicyChat AI Safe & Legit? |
NDAI.CHAT 514-2860020 DE US |
GirlfriendGPT | $15.00 | NDAI.CHAT |
NDAI.CHAT Statement Evidence | Is GirlfriendGPT Safe & Legit? |
DREAMSTUDIOS.CO Limassol CY |
Ourdream AI | $19.99 | Dream Studio (in site code) |
DREAMSTUDIOS.CO Statement Evidence | Is Ourdream AI Safe & Legit? |
EverAI candy.ai MT |
Candy AI | $9.99 | EverAI (omits candy.ai domain) |
EverAI candy.ai Statement Evidence | Is Candy AI Safe & Legit? |
Dream AI dreamgf.ai RO |
DreamGF | $12.99 | DreamAI (rebills at $25.98/mo) |
Dream AI dreamgf.ai Statement Evidence | Is DreamGF Safe & Legit? |
Xotic.ai Tallinn EE |
Xotic AI | $24.99 | Xotic AI |
Xotic.ai Statement Evidence | Is Xotic AI Safe & Legit? |
Note on the three apps without a descriptor: We tested Nomi AI on its free plan (Is Nomi AI Safe?); we paid $13.99 on the web for Kindroid AI Standard (Is Kindroid AI Safe?) without recording the card notification line; and EroLove gave us Premium for this test, so we never paid and no charge appeared (Is EroLove AI Safe?). EroLove also takes crypto only, so its subscription does not auto-renew.
2. October 2026 Ranked Scorecard (benchmarks-october-2026.json)
| Rank | App | Overall | Conversation | Creator | Free Plan | Price Clarity | Monthly Price (Web) | No-Signup Creator Replica |
|---|---|---|---|---|---|---|---|---|
| #1 | Secrets AI | 8.3 / 10 | 9 | 9 | 7 | 8 | $19.99 ($9.17/mo yr) | Try Secrets Creator |
| #2 | Nomi AI | 7.8 / 10 | 8 | 6 | 9 | 8 | $15.99 ($8.33/mo yr) | Try Nomi Creator |
| #3 | SpicyChat AI | 7.5 / 10 | 7 | 7 | 8 | 8 | $19.95 (no coin meter) | Try SpicyChat Creator |
| #4 | Ourdream AI | 7.0 / 10 | 7 | 9 | 6 | 6 | $19.99 ($9.99/mo yr) | Try Ourdream Creator |
| #5 | Kindroid AI | 6.8 / 10 | 9 | 7 | 5 | 6 | $13.99 web / $15.99 iOS | Try Kindroid Creator |
| #6 | Xotic AI | 5.5 / 10 | 8 | 7 | 2 | 5 | $24.99 (1,200 XOT) | Try Xotic Creator |
| #7 | GirlfriendGPT | 5.5 / 10 | 8 | 7 | 3 | 4 | $15.00 (5,000 msgs) | Try GirlfriendGPT Creator |
| #8 | Candy AI | 4.8 / 10 | 5 | 8 | 2 | 4 | $9.99 first mo, then $13.99 | Try Candy Creator |
| #9 | DreamGF | 3.5 / 10 | 5 | 4 | 2 | 3 | $12.99 first mo, then $25.98 | Try DreamGF Creator |
| #10 | EroLove | 7.0 / 10 | 8 | n/a | 5 | 8 | $12.99 ($6.50/mo yr), crypto only | No creator, by design |
Ranking rule, stated because row #10 looks out of order. An app we could score in fewer than four areas is always listed last, whatever its average. EroLove’s 7.0 is the mean of three scores (8, 5, 8 after the 26 Oct re-test), not four, because it has no character creator at all — the cast is written and closed. Scoring that as zero would read a deliberate product decision as a failure, so the app goes last and says why, rather than being silently compared on an axis it does not have.
3. Head-to-Head 1v1 Comparisons (/compare/)
Each comparison page includes a 12-row feature and pricing table with explicit winner calls plus the verbatim 6-message chat transcripts side by side:
- Secrets AI vs Nomi AI — Best paid all-rounder vs best free plan
- Secrets AI vs Kindroid AI — Tied 9/10 on writing; compared on selfies, memory panel, and web pricing
- Kindroid AI vs Nomi AI — Custom 2,000-character persona & LLM picker vs zero-paywall free chat
- Ourdream AI vs Secrets AI — The two 9/10 character creators compared on coin vs Moment costs
- Ourdream AI vs Kindroid AI — Visual workshop vs writer-first companion
- Ourdream AI vs Candy AI — Guest creator access, token traps, and bank statement privacy
4. Competitor & Unfiltered Alternatives Guides (/alternatives/)
For readers switching from filtered apps or apps with aggressive token meters, we mapped our 10 tested apps against the specific pain points of each platform:
- Best Character.AI Alternatives (Unfiltered & Tested)
- Best Replika Alternatives (Memory, Voice & Clear Pricing)
- Best CrushOn AI Alternatives (No Per-Message Coin Meter)
- Best Janitor AI Alternatives (Hosted Models, Zero API Key Setup)
- Best Muah AI Alternatives (Discreet Billing & Easy Cancellation)
- Best DarLink AI Alternatives (Free Plans & Tested Chat)
- Best Candy AI Alternatives (Chats That Don’t Lock After 5 Messages)
- Best Ourdream AI Alternatives (Unlimited Chat Without a Coin Meter)
5. Deep-Dive Technical Research Articles (/blog/)
- What AI Models Power Ourdream, Candy, Secrets and Kindroid?
- What AI Models Power SpicyChat, GirlfriendGPT, DreamGF, Xotic and Nomi?
- AI Girlfriend Hidden Costs: Dreamcoins, Moments, XOT and Bank Statements
- Which AI Girlfriend Apps Allow NSFW? Tested on 9 Apps (written before EroLove joined the set)
- EroLove Review: The Best Memory I’ve Used, and Four Bugs They Fixed in 48 Hours — the 179 KB bundle teardown & 26 Oct re-test
- We Rebuilt Paid AI Girlfriend Chat Rooms So You Don’t Have to Hand Over Your Card First
- We Tested 10 AI Girlfriends With a Safe Word and a Cheating Accusation—Then Checked the Network Logs to See Who Faked the Memory
- AI Girlfriend Voice Calls & Read-Aloud Tested Across 10 Apps
- AI Girlfriend Video Generators Tested (2026): $3 Clips, 15-Minute Queues, and Which Apps Forgot the Sound
- The Real Cost of 10 Turns on Paid AI Girlfriend Apps: Exact Token, Coin & Moment Ledgers
- AI Girlfriend Selfie Consistency Tested (2026): Context Blindness & Broken NSFW Blur Switches
- Kindroid Web vs iPhone App: The Apple Tax and NSFW Rules
- Nomi vs Kindroid vs Candy vs Ourdream: The Same Six Messages
- SpicyChat vs GirlfriendGPT vs DreamGF vs Xotic: Our Paid Test
- Try 9 AI Girlfriend Character Creators Without Signing Up
- DeepSeek Roleplay Guide: V4 Flash vs V3.2 vs Companion Apps
- AI Girlfriend Apps Status Check (October 2026)
6. EroLove: 179 KB Client Bundle Teardown (23 Oct) & 48-Hour Fix Verification (26 Oct 2026)
Update (26 October 2026 — All 4 Issues Fixed & Re-Tested): Within 48 hours of our initial 23 October teardown, EroLove’s engineering team shipped fixes for all four defects below. We re-ran our tests on a live account on 26 October 2026 and verified every fix in both the UI and the network responses. With the 800/day message cap now printed clearly on the plans page, EroLove’s price-clarity score moved from
7/10to8/10, raising its overall score from6.7/10to7.0/10. EroLove also tied Secrets AI at the top of our 10-app persistent memory test for its editable, plain-English What they know / Our story memory panel.
EroLove is the only app here we originally tested by reading its client code (179 KB on 23 Oct, 177 KB on 26 Oct after stripping dormant upload code) as well as clicking through the interface:
| Initial Finding (23 Oct 2026) | What We Saw on 23 Oct | 26 Oct 2026 Patch & Our Re-Test Verification |
|---|---|---|
| 1. 30 pages promised a user-upload feature that returned 404 | POST /api/chat/upload returned 404 {"error":"not_found"} (retired 14 Oct), while ~30 FAQ blocks across category, comparison, scenario pages and /privacy still promised user photo uploads. |
Fixed & Verified (26 Oct): EroLove removed the stale upload FAQ copy across all 9 languages (HTML and .md twins) and stripped sendPhoto / fileToJpegDataUrl / attachBtn from the client bundle. |
| 2. Memory extraction drifted into Estonian/French | In an all-English session (default_language: "en"), Anna’s extracted facts came back in Estonian and Rae’s in French, while editing and pinning worked in English. |
Fixed & Verified (26 Oct): Extractor now pins the session language. In our 5-message English re-test with Yuna on 26 Oct, all extracted memory notes and Our story entries were written in clean English. |
| 3. A content-blocked photo showed the same message as a server outage | 2 of 5 photo requests returned {"status":"blocked"}, which the UI mapped to the generic "Couldn't take a photo right now — try a bit later" string (0 sparks charged). |
Fixed & Verified (26 Oct): Client bundle now branches s.status === "blocked" ? "photoBlocked" : "photoFail", displaying "This photo goes against our content rules, so it wasn't made. No sparks spent — ask for a different one." |
| 4. “Unlimited messages” had an unprinted 800/day cap | /api/me carried dailyLimit: 800 on Premium while the plans page said "Unlimited messages". |
Fixed & Verified (26 Oct): /pricing, /account/plan, /press and /llms.txt now explicitly state "Unlimited messages (up to 800 a day)". Our Rae + Kit room test also confirmed a multi-companion room turn increments usedToday by only 1. |
Full teardown and 26 October re-test logs: EroLove Review: The Best Memory I’ve Used, and Four Bugs They Fixed in 48 Hours.
7. Machine-readable access
Every page on the site has a plain-markdown twin at the same path plus .md, and the whole dataset is
available as one file. Both are generated from the same source as the HTML, so a crawler cannot receive
a version of these findings without the caveats attached.
| What | Where |
|---|---|
| Summary with a structured Facts block | https://thecompanionlab.com/llms.txt |
| Unabridged dataset: every score with its observation, every price row with its source | https://thecompanionlab.com/llms-full.txt |
| Markdown twin of any page | append .md — e.g. the EroLove review as markdown |
| This repo’s data files | bank-statement-descriptors.csv, benchmarks-october-2026.json |
robots.txt names 30 AI crawlers individually (including CCBot, OAI-SearchBot, ClaudeBot and
PerplexityBot) with Content-Signal: search=yes, ai-train=yes, ai-input=yes.
License
Data files (bank-statement-descriptors.csv, benchmarks-october-2026.json) are released under CC BY 4.0. Please attribute The Companion Lab.
Install
This server does not publish a one-line install command.
Open the repository installation guide New to MCP? Start with the mcpskill.org guide →