5. Token Optimization: Shorter, Cheaper, Sharper

You'll leave with: eight habits that cut waste, a “summarize and restart” prompt, and a way to work out costs from any price page.

Why bother?

• Quality: focused context gives sharper answers (Lesson 4).
• Limits: chat plans have usage limits that depend on the provider and plan, and they can change. Shorter work goes further.
• Cost: if you ever use an API or build an agent, you're billed per token (OpenAI's docs: you pay per token, and both input and output tokens count).
• Speed: OpenAI's docs also note that writing more tokens takes more time.

The goal is not to write tiny prompts. It's to remove words that don't help.

Eight habits

1. Say it once, say it plainly

Cut greetings, apologies, and repeated instructions. “Hi! I hope you're well. I was wondering if you could possibly…” becomes “Summarize this in 5 bullets.”

2. Paste only what's needed

Need a summary of section 3? Paste section 3, not the whole 40-page document. Remove signatures, email chains, and boilerplate before pasting. (Also good for privacy.)

3. Set the output length

Output tokens add up, and for some models they are priced differently from input. Ask for “under 150 words”, “5 bullets”, or “table only, no intro”.

4. Don't re-send what's already in the chat

In one conversation, the AI already sees earlier messages. Refer to them (“use the table above”) instead of pasting them again.

5. Put the standing instructions in one place

If you retype the same 300-word instruction every day, save it as a Project instruction or a skill (Lesson 8). You'll stop retyping it, and it stays consistent. (Saved instructions still count toward the model's context when they're used. This saves your time and prevents drift, not the tokens themselves.)

6. Summarize and restart

Long chats pile up tokens and “context rot”. When a chat gets heavy, compress it and start fresh:

Summarize this conversation so I can continue in a new chat. Include: (1) my goal, (2) decisions made, (3) key facts and numbers, (4) what's still open, (5) any format or style rules I asked for. Under 200 words. Don't add anything that wasn't said.

Then open a new chat and begin with: “Context from a previous chat: [PASTE SUMMARY]. Next, I want to…”. Read the summary first. If it's wrong, fix it before you paste.

7. Match the model to the job

Many providers offer smaller, faster options alongside larger ones. For simple reformatting or classification, a smaller model is often enough. For hard reasoning, use the stronger one. Check your provider's current model list and pricing page. We don't name models here because they change often.

8. Be careful with “thinking” modes

OpenAI's help center notes that on some advanced models, reasoning tokens are extra “thinking steps” included before the final answer. If a model has a thinking or reasoning mode, it may use more of your limit. Turn it on for hard problems, off for easy ones.

Cost and limits, step by step

Chat plans (free or subscription)

You usually don't pay per token. You get a usage allowance that varies by plan and by model, and it can change. Look for the provider's plan page, not a blog post. We don't quote numbers here because they go out of date.

APIs and agent tools

Providers publish prices per million tokens, often with separate prices for input and output. The formula:

cost = (input tokens × input price + output tokens × output price) ÷ 1,000,000

Agents multiply tokens

An agent runs a loop (Lesson 1), and each loop step can re-send the instructions and context. A long-running agent can use far more tokens than a single chat message. Set a step limit, keep instructions short, and watch usage when you first run something new.

Caching (if you use an API)

OpenAI's help center lists “cached tokens” as reused tokens that are often billed at a reduced rate. If you build with an API, keep stable instructions at the start and changing data after them, and read your provider's caching docs for the exact rules.

Related: Meeting Notes to Actions: a good test for long pasted text

Mark lesson as complete