Tokens, and how to manage what you spend on them
Tokens get cheaper every year, and most bills still go up. Here is what a token is, what drives the cost, and the handful of settings that control it.
This page is general guidance from working with these tools, not a cited source. Prices and plan limits change often, so check each provider's pricing page before you budget.
1. What a token is
A token is a chunk of text, roughly three quarters of an English word. Numbers, code, and other languages often use more.
About 1,000 tokens is roughly 750 words, or a page and a half of text. A 40-page quarterly report can be 25,000 tokens or more.
You pay for tokens going in (your prompt, files, and the conversation so far) and tokens coming out (the answer). Output usually costs several times more per token.
2. Subscriptions vs. pay-per-token
Chat subscriptions
ChatGPT, Claude, and Gemini business plans charge per person per month. You don't see tokens directly, but they still matter: heavy use hits usage limits, and the top tiers exist because some people burn through more.
Best for: most staff and family members doing everyday work.
API and agents
When you connect AI to your own tools, run agents, or use coding assistants on usage-based billing, you pay per token. This is where bills surprise people, because one agent task can take dozens of steps and re-read the same material each time.
Best for: automated workflows, agents, and tools you build.
3. What drives the bill
Four things explain most of the cost.
Which model
The largest, most capable models can cost 5 to 30 times more per token than the small, fast ones from the same lab. Many routine tasks don't need the biggest model.
How hard it thinks
"Reasoning" or "thinking" models work through a problem before answering, and those thinking tokens are billed like output. A higher effort setting can multiply the cost of one answer.
How much it reads
Every turn in a conversation re-sends the conversation so far. A long chat with three attached reports gets more expensive with each question, even short ones.
How many steps
An agent that searches, reads, drafts, checks, and revises may make 20 to 50 calls for one task. Each call pays for what it reads and writes.
4. Ten ways to manage token spend
| Lever | What to do |
|---|---|
| Match the model to the job | Use a small, fast model (the "mini," "flash," or "haiku" tier) for sorting, extraction, and short summaries. Use the mid tier for most drafting and analysis. Save the frontier model for hard reasoning, important memos, and code. |
| Set effort on purpose | Where a tool lets you choose effort or thinking level, default to low or medium. Turn it up for a hard problem, then turn it back down. |
| Start a new chat for a new topic | Long threads re-send everything every turn. When the subject changes, start fresh and paste a short summary instead. |
| Send the pages, not the binder | Attach the section that matters rather than the whole 200-page document. A notebook tool like Gemini Notebook is built to search large source sets for you. |
| Use skills and projects | Store your instructions once as a skill or project rather than pasting a long prompt every time. It's cheaper, and the result is more consistent. |
| Ask for shorter answers | Output costs the most. "One page," "five bullets," or "table only" cuts both cost and reading time. |
| Use caching for repeated material | On the API, prompt caching lets the model re-read the same long instructions or documents at a steep discount. Ask whoever builds your tools to turn it on. |
| Batch work that can wait | Batch processing on the major APIs runs non-urgent jobs (overnight summaries, document intake) at roughly half the normal price. |
| Cap agents | Give each agent a step limit and a spending limit per task. An agent stuck in a loop is the most common surprise bill. |
| Meter and review monthly | Set a monthly cap and an alert at 50% and 80% in each provider's console. Review spend by workflow, not just by person, and compare it with the number the workflow was meant to move. |
Budget for AI like a utility: cap it, meter it, review it monthly. And measure cost per outcome, such as cost per meeting brief or per capital call processed, not cost per token.
5. Estimate a workflow's monthly cost
Plug in your provider's current prices. The defaults are placeholders for illustration, not quotes.
Try it: drop the model calls from 12 to 3 (a simpler workflow), or the input tokens from 20,000 to 5,000 (send the pages, not the binder), and watch the monthly number.
6. Why the bill still goes up
cheaper per token since 2023 (Apollo and Fortune, June 2026, citing Silicon Data)
more tokens a month at Google, in one year (Google I/O, May 19, 2026)
Jevons paradox, 1865
When steam engines got more efficient, Britain burned more coal. Cheaper tokens mean more uses, more agents, and bigger bills.