How to use WhatModel

Part 1 — Concepts (read once)

What WhatModel is

AI models (like GPT-4o, Claude, Llama, Gemini) are sold on a pay-per-use basis, priced per million words of text you send in or get back. Prices, quality, and availability shift constantly — and different companies offer the exact same model at different prices. WhatModel watches all of that and shows you where the market is right now.

Creator vs Host

Creators are the labs that build a model — OpenAI, Anthropic, Meta, Google, Mistral, DeepSeek, and so on. Hosts are the companies that run the model on their servers and rent it out to you. A model can have one host (only its creator serves it, like OpenAI serving GPT-4o) or many hosts (Meta's Llama can be run by Together, Fireworks, Groq, DeepInfra, and dozens more). Same model, different prices, different speeds. That's where opportunity lives.

Tokens, context, input & output prices

A token is roughly ¾ of a word. When you use an AI model, you pay for the tokens you send in (your prompt) and the tokens it sends back (its answer). Prices are usually quoted per million tokens — for example "$3 input / $15 output" means you'd pay $3 for every 1M words you send and $15 for every 1M words you get back.

Context window is how much text the model can hold in its head at once — like short-term memory. 8K context means it can consider about 6,000 words at a time; 1M context means it can read a whole novel. Bigger context = more expensive to run but more useful for long documents.

Quality benchmarks

Models are graded by independent leaderboards: LMArena (human head-to-head voting), MTEB (embeddings), and Artificial Analysis (mixed tests). WhatModel pulls those scores in as-is. A higher score means the model wins more head-to-head comparisons — but it isn't the whole story. Cheap-but-good beats expensive-and-great when the job is simple.

WM Score

WM Score is our 0–100 composite rating. It blends five signals: price, context window, capabilities (tool use, vision, function calling), uptime, and host redundancy (how many places you can run it). We keep the exact weights private, but higher is better across "does the job cheaply, reliably, and everywhere." The score is recomputed every 12 hours and only fires a change event when the move is meaningful (≥5 points), so the feed stays quiet unless something real happened.

Spread %

Spread is the gap between the cheapest and most expensive host for the same model, expressed as a percentage. A 40% spread on Llama-3.3 means one host charges 40% more than another for the identical model. Spread is our #1 signal that you might be overpaying. It's not "arbitrage" you can execute automatically — WhatModel doesn't route your traffic — but it tells you where to move.

Uptime, latency, throughput

Uptime: what % of requests succeeded over the last 30 minutes. Latency: how long you wait for the first word (measured in ms). Throughput: how many words per second the model streams to you. Cheap hosts sometimes trade one of these for a lower price. Fast hosts sometimes cost more. We show all three so you can pick your trade-off.

GPU spot market

Some users run models themselves on rented GPUs (H100, A100, RTX 4090). Those GPU rentals also have a live spot price — companies like Vast.ai and RunPod publish it. We track that too. If GPU rent drops below the cost of paying a host, self-hosting becomes cheaper. That gap is what the /arbitrage page will show once we've collected enough history.

Watchlist, alerts & webhooks

Star any model to add it to your watchlist. When a price moves, a new host appears, a capability is added, or quality drops, we fire an alert. Alerts show up in the bell menu; Pro users can also send them to a webhook (a URL we POST to) for Slack, Discord, or automated pipelines.

Part 2 — Page-by-page

Part 3 — Common recipes

"I want the cheapest host for GPT-4o"

Open /search, type "gpt-4o", click the row. In the model detail page, sort Host Offerings by "Input $" ascending — the top row is your answer. Check uptime and latency before committing.

"I want to know when Claude's price drops"

Open the Claude model page, click the ⭐ star to add it to your watchlist. When any host drops its price, an alert fires. Pro users can add a threshold (e.g. only tell me if the drop is >10%) or route it to a webhook.

"I want to compare two models"

On any model card, click the compare-tray icon. Do it again on the second model. Open /compare. Free users get 2 slots; Pro gets 4.

"I want to know if self-hosting is cheaper"

Open /router. Enter your monthly input and output token estimate. The tool shows the cheapest host mix and the break-even against renting a GPU yourself.

"I want a weekly summary"

Bookmark /digest for the weekly market snapshot.

Part 4 — Reference

Keyboard shortcuts

• ⌘K or Ctrl+K — command palette
• / — focus search
• S spot matrix · A arbitrage · H hosts · M models · R rankings
• G GPU market · O router · W watchlist
• ? — help modal

Free vs Pro

Glossary

Token — ≈¾ of a word. The unit of AI billing.

Context window — How much text the model can consider at once.

Input price — Cost per 1M tokens you send in.

Output price — Cost per 1M tokens the model sends back.

Creator — The lab that made the model.

Host — A company that runs the model and sells access.

Spread % — Price gap between cheapest and most expensive host for the same model.

Uptime — % of requests that succeeded in a window.

Latency — Time to first token (ms).

Throughput — Tokens per second while streaming.

WM Score — 0–100 composite: price, context, capabilities, uptime, redundancy.

Quality — Independent leaderboard score (LMArena, MTEB, etc.).

Arbitrage delta — Gap between cheapest hosted price and self-host cost on rented GPU.

Webhook — A URL we POST to when your alert fires.

Sparkline — Tiny inline chart showing 7-day price trend.

FAQ