☥ CERES · THE ANKH STATION NOUS · OPENROUTER · GOOGLE

THE GUY 520

Plain-English pocket guide to all 520 live AI models, broken down from Level 0 (Free) to Level 5 (Apex Expensive), with Owner Symbols, Curiosities, Retired Legends, and The Gare task router.

1. How the 6 Rudimentary Levels Work (0 to 5)

Every model across Nous, OpenRouter, and Google is sorted by what 1 million tokens actually cost:

Level 0 · Free ($0.00 / $0.00) — 23 Models
Costs zero dollars. Includes the 4 Nous Portal free models (Solar, Step, Beam, Ornith) + 17 OpenRouter free models (Inkling 975B, Nemotron Ultra 550B, North Code).
☥ Circle 0: Shrine of Humility · Reagent: Pure Spring Water
Level 1 · Almost Free (≤$0.25 In / ≤$0.60 Out) — 144 Models
Pennies per million tokens. Home to high-speed Flash/Mini workhorses (Gemini 2.5 Flash Lite, Claude Haiku 5.5, GPT-6 Luna, Qwen 3.8 Flash, Mercury 2.5 Diffusion) + 34 Vector Embedders.
☥ Circle I: Apprentice Cantrips · Reagent: Sulfurous Ash & Garlic
Level 2 · Economy Workhorse (≤$0.85 In / ≤$2.50 Out) — 133 Models
Strong daily drivers under $2.50/M output: DeepSeek V3.2 & R1, Gemini 2.5 Flash, GPT-5 Mini, Qwen 3.5/3.6 Plus, Kimi K2.5.
☥ Circle II: Journeyman Wayfarers · Reagent: Ginseng & Blood Moss
Level 3 · Mid-Tier Specialist (≤$2.50 In / ≤$8.00 Out) — 107 Models
Fast flagship-adjacent brains: Gemini 3.8 Flash ($0.82/$4.12), Grok 4.7 ($2/$6), Qwen 3.8 2.4-Trillion ($2/$6), GPT-6.1 Terra ($1/$5), Hermes 4 405B ($1/$3).
☥ Circle III: Adept Paladins · Reagent: Spider Silk & Black Pearl
Level 4 · High Frontier (≤$6.00 In / ≤$25.00 Out) — 87 Models
The primary heavyweights: Gemini 4 Argon ($2/$10), Claude Sonnet 5.5 ($2/$10), GPT-6.1 Sol ($2/$10), Claude Opus 5.5 ($4/$20), GPT-5.4 ($2.50/$15).
☥ Circle IV: High Magi of Britannia · Reagent: Mandrake Root
Level 5 · Apex Expensive Frontier ($7.50–$30 In / $30–$180 Out) — 26 Models
The Royal Treasury burners: Claude Fable 5.1 ($10/$50), GPT-6 Astra Pro ($10/$50, or $20/$100 Fast), o3-pro ($20/$80), and GPT-5.5 Pro ($30/$180).
☥ Circle V: Codex of Ultimate Wisdom · Reagent: Nightshade

2. Plain-English Decoder: What the Model Words Mean

MoE (e.g. "501B-A23B"): Mixture-of-Experts. The model stores 501 Billion parameters of knowledge, but only fires 23 Billion per word—giving you giant-model brains at small-model speed.
Dense (e.g. "122B" or "31B"): Every single parameter fires on every word. Slower per parameter, but often thicker and more consistent on subtle prose and translation.
Flash / Bolt / Lightning: Tuned for high speed and tool-calling loops where waiting 30 seconds kills momentum.
Lite / Nano / Mini: Distilled smaller models built for cheap classification, tagging, and simple JSON chores.
:batch (36 Models): Half-price (50% off) asynchronous lane for non-urgent jobs.
~latest (Rolling Pointer): Starts with a tilde (~openai/gpt-astra-latest); automatically floats to the newest model in that family.

🚉 3. "The Gare" — Which Model to Call for Which Task

Pick the right champion for the job at $0 Free, Level 1 Cheap, or Level 4 Frontier:

🗺️ Realm Map of the 65 Model Creators

Tap any symbol on the map or tap a House card below to jump straight to all models built by that creator:

⚡ The Anomaly Museum (Weird Architectures & Price Traps)

Not every model is a normal chatbot. Watch out for these 7 major anomalies across the 520 models:

1. The 1,133x Price Ratio Trap (z-ai/glm-5.2): Input is $0.006/M, Output is $6.80/M (1,133x pricier to speak than to read!). Similarly, kimi-k3 is $0.19 in / $15 out (78x).
2. The Lyria Audio Billing Trap (lyria-3-clip/pro): Shows $0.00 token price in JSON, but charges $0.04 per 30s music clip or $0.08 per full song.
3. Diffusion Text LLMs (inception/mercury-2.5 & mercury-3): Generates whole blocks of text simultaneously via diffusion instead of word-by-word (~1,000+ tok/s).
4. 1.58-Bit Ternary Weights (prism-ml/ternary-bonsai-2-27b): Every neural weight is strictly -1, 0, or +1. Zero floating-point math.
5. Hybrid Mamba-Transformers (nvidia/nemotron-3-ultra-550b-a55b): Fuses state-space Mamba with Transformer layers for 1M context at 55B active parameters.
6. Fast-Apply Code Mergers (relace-apply-3, morph-v3): 10,000 tok/s patch appliers meant to merge code edits, not chat.
7. HTML-to-JSON Scrapers (schematron-v2): Built solely to turn messy web HTML into strict JSON schemas ($0.03/$0.15).

🪦 The Tomb of Legends (Laguna, Space-Bunny & Retired Ghosts)

Models that used to exist (or still haunt static menus) but now return 404 or closed their free doors—and what to use instead today:

🔮 "The Tell" — 8-Model × 5-Field Combinatorial Matcher

Tap any combination of the 5 fields below to dynamically rank which free models respond accurately and fastest for that exact mix of skills: