// AI companion platforms, quantitative benchmark
// 12 platforms tested, 5 surviving the methodology cut.
// Composite weights: persona-consistency 0.25, visual-fidelity 0.20,
// latency-p95 0.15, conversation-depth 0.15, pricing-clarity 0.10,
// account-deletion 0.10, mobile-parity 0.05.

apg@bench:~$ ./run-benchmark --scope=all --weights=composite-v2
[done] 12 platforms, 5 ranked, 7 below floor.

AI companion platforms, ranked by composite benchmark.

Benchmark-driven reviews of AI companion platforms. Less vibes, more logs.

Ranked composite benchmark, May 2026

↓ click any row · cycle refresh monthly

# platform composite score visual starting / mo free tier strongest dimension
01 secrets-ai 9.2 From $9.99/mo voice
02 funfun-ai 9.1 From $12.99/mo image-generation
03 muah-ai 8.9 From $29/mo nsfw
04 spicychat 8.8 From $14.99/mo roleplay
05 swipey-ai 8.4 Free trial + premium voice
06 joyland-ai 8.4 Free + paid sustained-conversation
07 dreamgf 8.3 From $12.99/mo image-generation
08 janitor-ai 8.2 Free trial + premium roleplay
09 girlfriendgpt 7.9 From $14.99/mo roleplay
10 thotchat-ai 7.9 From $14.99/mo nsfw
11 juicychat-ai 7.8 From $29/mo roleplay
12 joi-ai 7.8 From $29/mo nsfw

Composite is the weighted geometric mean of nine sub-scores. Sub-scores at /method.

Cohort index

$ ls -al ./benchmarks/

# cohort scope
01 best-ai-girlfriends full composite ranking
02 ai-girlfriend sustained conversation cohort
03 best-nsfw-ai-chatbots uncensored cohort
04 ai-anime anime-first cohort
05 best-ai-girlfriends-with-voice voice latency cohort
06 best-ai-girlfriends-with-images image-gen cohort
07 best-roleplay-ai-chatbots long-arc roleplay cohort
08 best-free-ai-girlfriends free-tier cohort
09 ai-characters character marketplaces
10 sexting-ai nsfw + roleplay overlap
11 ai-boyfriend male persona configurable

$ open /aitool/ for the full platform index

Sub-score heat-strip, strongest dimension per platform

Each row is a platform. Each cell is one of the nine sub-scores, on a 0 to 10 scale. Darker = higher. Pick the platform that’s strongest on the axis you care about, not the one that wins the composite.

persona-cons.
visual-fid.
latency-p95
conv-depth
pricing-clear
acct-del.
mobile-par.
boundary
moder-cons.
secrets-ai
9.7
8.4
10.0
8.9
9.1
9.5
9.6
9.2
8.6
funfun-ai
9.8
9.6
10.0
8.8
9.1
9.1
9.6
8.3
8.6
muah-ai
9.1
8.8
9.8
9.2
9.7
9.4
8.4
8.1
8.2
spicychat
9.6
8.7
9.4
8.2
8.1
8.3
8.1
8.4
8.7
joyland-ai
8.2
8.7
9.0
8.5
8.9
7.5
8.6
7.6
8.8

Methodology log, May 2026 cycle

2026-05-02 · cycle start · 12 platforms shortlisted · subscriptions paid out-of-pocket · no contact with vendors.
2026-05-04 · editor-1 begins 8-hour sustained sessions · editor-2 begins multi-persona rotation.
2026-05-09 · latency probes from 3 PoPs (frankfurt, virginia, singapore) · p95 collected over 600 requests / platform.
2026-05-12 · account-deletion test on each platform · 4 platforms pass single-click bar · 3 fail dark-pattern check.
2026-05-17 · sub-scores reconciled · composite locked · pre-publication freeze for 48 h.
2026-05-19 · publish.
→ full methodology & rubric weights