Underlying AI Models & Roleplay Benchmarks
What actually runs under the hood when you pay $19.99/month for an AI companion? We cross-referenced our 217 reverse-engineered JavaScript bundles with live OpenRouter API tests, HuggingFace model cards, Civitai checkpoints, and EQ-Bench Slop scores.
Live OpenRouter Roleplay Arena & EQ-Bench Slop Detector
Lab Take: In 17 words, V3.2 invents a concrete sensory detail ("vanilla lattes") that catches you in a lie. Zero purple prose, zero wasted tokens—this is why dream-girl (梦女) roleplayers fell in love with V3.2.
Lab Take: When we capped completion at 220 tokens, V4.1 Flash returned a blank string because it burned all 220 tokens thinking to itself! Uncapped, it burned 425 reasoning tokens ($0.00058 vs $0.00005) to write a decent but choppy screenplay beat.
Lab Take: Reads like a daytime soap opera script mixed with a therapy worksheet ("You do realize how that might make me feel"). Zero subtext.
Interactive EQ-Bench "AI Slop & Cliché" Detector (Paste Any Companion Reply)
Based on eqbench.com/slop-score.html trigram & contrast forensicsAudited Model & Checkpoint Dossiers
Why haven’t AI companion apps upgraded from DeepSeek V3.2 to the newer DeepSeek V4.1 Flash? Our live OpenRouter API benchmark proved why: V4.1 Flash is tuned for agentic coding and tool-use, burning 425–512 hidden reasoning tokens per message only to sound like a cautious HR manager.
Discovered inside Ourdream.ai’s production bundle under the enum keys ZImage and ZImageMoody, Tongyi Z-Image-Turbo is the reason Ourdream renders realistic skin pores and bedroom lighting in 4 seconds flat.
The gold-standard anime and 2.5D character checkpoint on Civitai, wrapped inside Ourdream AI’s "Fluidity" mode for instant browser generation.
Our REA bundle extraction caught both Secrets.ai and Ourdream.ai calling ByteDance’s Seedream 5.0 Pro and Alibaba’s Wan video APIs—explaining both their impressive motion quality and why 8 video clips drain a whole month of coins.
Built specifically to eliminate the robotic "As an AI, I cannot..." lecture while keeping character voices grounded, Euryale 70B and Hermes 4 405B represent the benchmark for uncensored adult storytelling.
ELI5 Encyclopedia: AI Companion Terms Explained Like You’re 5
The maximum amount of active text (measured in tokens) an AI model can hold in working memory at one time before older messages drop off.
Image & Video CheckpointsTongyi Z-Image-Turbo (S3-DiT) →A 6-billion-parameter Single-Stream Diffusion Transformer (S3-DiT) developed by Tongyi-MAI that renders photorealistic images in 8 steps.
Image & Video CheckpointsWAI-Illustrious-SDXL →The #1 community-trained anime and 2.5D illustration checkpoint on Civitai (1.55M+ downloads), built on the Illustrious-XL architecture.
Writing & Slop BenchmarksEQ-Bench Slop Score →A quantitative metric created by EQ-Bench that measures how frequently an AI model repeats overused romance clichés, purple prose, and "Not X, but Y" rhetorical tics.
Billing & PrivacyApple IAP Tax →The 14% to 20% price markup users pay when subscribing to an AI companion inside the iOS App Store or Google Play instead of the official web checkout.
Image & Video CheckpointsSeedream 5.0 Pro & Wan 2.5 →Commercial frontier image-editing (ByteDance Seedream 5.0 Pro) and image-to-video (Alibaba Wan 2.5 / Kling v2.5) APIs used behind the scenes by top AI companion studios.
Writing & Slop BenchmarksAgentic CoT Drift →The phenomenon where newer frontier models tuned for coding, tool-use, and reasoning (like DeepSeek V4.1 Flash) burn 70%+ of their tokens thinking like a project manager—making their romantic roleplay stiffer and 5x more expensive.
Billing & PrivacyBilling Descriptor →The exact company name and text string that appears on your monthly credit card or bank statement after you subscribe to an AI companion app.