Token Economics for Organizations

Subscriptions, APIs, and enterprise agreements all meter the same underlying thing - tokens. Learn how AI pricing really works so you can predict, compare, and negotiate.

The Three Ways Organizations Buy AI

1. Seat-Based Subscriptions

A chat assistant or coding copilot priced per user per month. Predictable and simple - but you pay for every seat whether it is used or not, and heavy users hit rate limits. When organizations "cancel an AI subscription," it is usually because seat utilization could not justify the price: 1,000 seats × $30/month is $360,000 a year, and utilization audits often find fewer than 40% of seats in weekly active use.

2. API Consumption (Pay Per Token)

You pay for exactly what you use: input tokens (everything you send - system prompt, conversation history, documents) plus output tokens (everything the model generates). This is where optimization has direct, immediate impact: every token you eliminate is money back, on every single call.

3. Enterprise Agreements

Committed-spend contracts blending the two, often with volume discounts, provisioned throughput, or private deployments. Your negotiating position is entirely determined by how well you understand your own token consumption profile - which is exactly what Lesson 7's measurement practices give you.

How API Pricing Works

Prices are quoted per million tokens (MTok), with output typically costing 3-5× input. The tiers below are illustrative - always check current vendor pricing, because these numbers change frequently:

Context Windows Are Priced Linearly - But Used Carelessly

A 200k-token context window is a capability, not a goal. Every token in the window is billed on every call. Teams that "just stuff everything into context" pay for that convenience on every single request - and pay again as conversations continue and that context is re-sent turn after turn. Lesson 4 covers the discipline of deciding what actually deserves a place in the window.

Total Cost of Ownership: Tokens Are Not the Whole Story

A fair comparison between models also counts:

• Retry and failure tokens. A cheaper model that fails 20% of tasks and needs retries can cost more end-to-end than a pricier model that succeeds first time - and that is before counting the downstream cost of wrong answers that slip through.
• Engineering time. Heroic prompt gymnastics to make a small model work has a salary cost that rarely appears on the AI dashboard.
• Latency. Sometimes paying more per token but finishing in one shot is the cheaper system overall.

Subscription or API? A Quick Decision Frame

The organizations that get this right run the analysis with real usage data - per-seat activity for subscriptions, per-feature token logs for APIs - rather than gut feel. Lesson 7 shows how to have that data on hand permanently.

Check Current Pricing at the Source

The figures in this lesson are illustrative and will drift. When you build a cost model, pull current numbers from the vendors' official pricing pages:

• Anthropic pricing - Claude model tiers, prompt caching, and batch discounts
• OpenAI API pricing - GPT model tiers, cached input, and Batch API rates
• Google Vertex AI pricing - Gemini model tiers and provisioned throughput