A TOKEN is roughly 3/4 of a word. A page of text is about 500 tokens. This curriculum, when finished, will be ~30,000 tokens.

When you talk to a frontier model, you pay for:

• INPUT TOKENS: everything you sent the model (your prompt, attached files, conversation history).
• OUTPUT TOKENS: everything the model writes back.

Output tokens are typically 3-5x more expensive than input tokens. The exact numbers change by model and by month — check the provider's pricing page.

The CONTEXT WINDOW is the maximum number of tokens (input + output) the model can handle in one conversation. Frontier models in 2026 typically have 200K-1M token windows. Local models are usually smaller (8K-128K).

Why this matters for cost:

Long conversations cost more than short ones, BECAUSE EACH MESSAGE RE-SENDS THE WHOLE HISTORY. You're not just paying for the new turn — you're paying for the agent to "re-read" everything that came before. A 50-turn conversation costs ~50x what the first turn cost, on the last turn.

The fix:

START FRESH SESSIONS FOR NEW TASKS. When you switch to a different problem, start a new conversation. Carry over only what's needed.

KEEP SPECS IN FILES, NOT IN CHAT. The CLAUDE.md and spec docs in your repo are loaded on demand. They don't bloat every message.

USE LOCAL FOR ROUTINE WORK. If a task doesn't need frontier intelligence, run it locally. Even at $0.01 per call, the calls add up over hundreds of agent runs a day.

The $200 question: is Claude Max worth $200/month?

Math: $200/month is roughly 4-6 hours of a contract developer's time at $40-50/hour, or 1-2 hours at higher rates. If Claude saves you 4+ hours a month, it pays for itself. For most readers who are using it for serious work — building real software, generating real revenue — it pays for itself in the first week.

If you're learning, use the cheaper Pro tier first ($20/month, lower limits). Don't go to Max until you're hitting the Pro limits often enough that the upgrade math is obvious.

If $20/month is more than your hobby budget, run local. Quality is lower but real work is possible.

================================================================================ PART 5 — DESIGNING WHAT YOU'RE BUILDING