Anthropic's and OpenAI's most expensive models cost 100 times more per token than the smallest models in their lineups, yet many users default to the flagship for everything.

The cost of that habit is no longer hypothetical. Usage leaderboards at Meta, Amazon, and Uber pushed AI bills sharply higher, and Uber burned through its 2026 AI tools budget in roughly four months. Meta employees consumed 60 trillion AI tokens in 30 days earlier this year while chasing spots on an internal leaderboard nicknamed "Claudeonomics." Tech columnist Gergely Orosz estimated the tab at about $900 million at Anthropic's list prices, though Meta likely pays far less.

Bain \u0026 Company consultant Jue Wang sees a quieter version of the same waste in many organizations. "So many companies, so many users, default to using Opus for everything, including generating emails," she told the Associated Press this summer, referring to Anthropic's high-end models. Doing the same is akin to hiring Albert Einstein to calculate your change at McDonald's.

Every word typed into ChatGPT or Claude — and every word the model writes back — is metered in tokens, chunks of text that average about three-quarters of a word. On a $20 subscription, that meter determines how quickly the usage limit is reached. On the API, the pay-per-token access developers use to plug models into their own apps, it determines the bill.

Anthropic's current price list charges per million tokens, and Haiku's rate applies only to prompts under 100,000 tokens; longer prompts cost five times as much. The top model costs 100 times more than the smallest, and OpenAI's lineup runs the same spread, from $10 to $0.10 per million input tokens.

Those per-token prices add up fast for the labs — and for their customers. Anthropic's revenue run rate climbed from about $9 billion at the end of 2025 to more than $47 billion by May, while OpenAI fights a price war to keep up.

A few habits cut that spend sharply, and the first five tips below require no coding at all.

What tokenmaxxing means (and what it should mean)

Tokenmaxxing began as a Silicon Valley flex on the order of Meta's Claudeonomics, with Microsoft, Shopify, and Amazon ranking employees on similar token leaderboards. Nvidia CEO Jensen Huang poured fuel on the fire when he said on the "All-In" podcast in March that he would be "deeply alarmed" if a $500,000 engineer didn't burn through at least $250,000 worth of tokens.

And burn tokens they did — often on unproductive busywork. At Amazon, employees admitted inflating their numbers by running the internal AI tool on irrelevant tasks, until a senior vice president told staff: "Please don't use AI just for the sake of using AI."

This guide uses the meaning engineers are pushing now: maxxing the value of each token, not the number spent — the practice IBM describes as tokenmaxxing as a productivity metric.

Three concepts explain where the tokens actually go, starting with the agent loop. An agent is a model that can use tools, so it thinks, searches the web or runs code, reads the result, and goes again until it decides it is done — resending the entire conversation on every lap.

Orchestration multiplies that cost. One model acts as a manager, splitting a request into pieces, handing them to worker models called subagents, and stitching their answers back together. Each worker runs its own loop with its own memory, which is why Anthropic says agent teams in Claude Code can use about seven times more tokens than a standard session when teammates run in plan mode, a setting covered in the intermediate tips below.

Graphs are how developers keep those loops from wandering off. Frameworks like LangGraph let developers draw a workflow like a subway map, with stations (nodes) doing the work and tracks (edges) deciding where to go next, so the model follows a fixed route instead of looping until it runs out of ideas or budget.

Loops automate tasks and raise quality, but they also burn money — which makes efficiency essential.

Beginner: small jobs, small models

1. Pick the model for the job. Most beginners run AI through a graphical interface. Click the model name in the chat app (at the top of the chat in ChatGPT, next to the send button in Claude or Hermes) and switch to a smaller, faster option for emails, summaries, and rewording. Reserve the top model for hard reasoning and decisions where a subtle mistake costs real money. Run five of last week's tasks on both models side by side; if the small one passes, keep it. On a subscription, that stretches the usage limit, and on the API it cuts the bill by up to 99%.

2. Write one good prompt instead of five bad ones. Three rounds of "no, just the total" or "make this better and more accurate" resend the whole chat every time. One clear prompt spells out the job and the exact shape of the answer. Something like can work:

Extract the invoice number, date, vendor and total from the text below. Keep every value exactly as written. Return one line per field, and write "missing" if a field isn't there.

OpenAI's own testing found that leaner, clearer prompts scored higher while using fewer tokens.

3. Start a new chat when you change topics. The model rereads the entire conversation on every send, so a 50-message thread about a vacation makes a tax question pricier and dumber at once. Anthropic calls that decay "context rot". Start a new chat whenever the subject changes.

4. Ask for the size you need. Output tokens cost five times more than input on Anthropic's and OpenAI's top models. Ending prompts with instructions like "three bullets, under 100 words" or "only the changed lines" leaves the model less room to pad. Output length constraints save tokens, especially on creative tasks like writing or research rather than deterministic ones like coding or math.

5. Try free before you pay. OpenRouter, a site that connects to hundreds of AI models through one account, has a free router that sends requests to one of about 20 no-cost models. Models can also be run locally with LM Studio or Ollama on a graphics card with 8GB of video memory or an Apple M-series Mac. Free tiers can vanish overnight, so don't build anything important on one, and read the data policy before pasting anything sensitive. That said, free offers are a way to max out iterations: Hermes has a good selection of free models that may handle the job, and so does Nvidia. It is better to refine a prompt with a free, dumber model than to pay a premium model to do the same thing faster.

Additional beginner tips: skip extended thinking for simple questions, or lower the effort setting on newer Claude models where thinking can't be switched off — Anthropic bills the full thinking process as output, including the part users never see. Claude Pro and Max subscribers who turn on usage credits can set a monthly spend limit under Settings, then Usage.

Intermediate: your agent is reading way too much

The examples from here on use Claude Code, Anthropic's agent that runs in the terminal and edits files; commands go straight into its chat box. Other harnesses follow similar logic, and if only the chat apps are in play, tip 8 works in any chat window.

6. Start cheap, escalate on failure. Type /model to switch models, keep Sonnet as the default, and move up to Opus only when Sonnet fails at something specific, such as a bug it can't fix after two tries. For simple subagent jobs, Anthropic's docs suggest setting model: haiku in the subagent's settings.

7. Clear between tasks. Type /clear when switching to an unrelated task — it costs nothing and drops the old history. Use /compact with a note on what to keep, like /compact keep the database decisions, only when the history is genuinely needed, since compacting is itself a large request.

8. Let code do the boring math. Pasting a 5,000-row spreadsheet into a chat makes the model read every row, and it can still miscount; a language model is not a calculator at any price. Ask for a formula or a short script instead, run it locally, and hand the model only the result.

9. Trim your tools. MCP servers are plugins that connect a model to apps like GitHub or Google Drive, and each one adds tool names and instructions to every session. Type /context to see what is eating space and /mcp to disable servers the project doesn't use.

Additional intermediate tips: type /usage to see where the tokens went, and keep CLAUDE.md (or Soul.md) — the instruction file the agent reads at the start of every session — under 200 lines. For bigger changes, press Shift+Tab to switch to plan mode, where the agent proposes a plan for approval before touching files, and hit Esc the moment it heads the wrong way. Use /effort to dial reasoning down for simple jobs like renaming a variable. Anthropic showed an agent's token use dropping from 150,000 tokens to 2,000 when it loaded only the tool definitions it needed — a reason command-line tools like GitHub's gh, which cost nothing until they run, often beat an always-on plugin.

Advanced: build the guardrails into the system

Consider a developer who needs to process 1,000 invoices of 2,000 tokens each, with 200-token answers: the job costs about $30 on Fable 5.1 and about $0.30 on Haiku 5.5, before any of the tricks below.

10. Cache the parts that never change. Put the stable instructions first and the changing parts last, and mark the stable block for caching. In Anthropic's API, that is one extra field:

"system": [{
  "type": "text",
  "text": "Your invoice instructions...",
  "cache_control": {"type": "ephemeral"}
}]

Reading from cache costs a fraction of the normal input price, while writing to the default five-minute cache costs extra. Check the cache_read_input_tokens field in each response; if it still reads zero after the first call, the cache isn't working.

11. Batch anything that can wait. Both OpenAI and Anthropic process batch jobs at a 50% discount, with OpenAI promising results within 24 hours. Uploading the 1,000 invoices as one batch file instead of making 1,000 live calls drops the bill heavily.

12. Put agents on a leash. Set a maximum number of steps and a dollar cap for every agent run, and define what "done" looks like before it starts. Claude Code accepts a --max-budget-usd flag for scripted runs, and Shopify uses circuit breakers that can cut off a developer's access immediately if their spend spikes within a day, built to catch runaway agents. The current norm is to set agents free and let them loop until the result is deemed perfect; without the budget for that, don't let the agent decide what counts as "perfect."

Additional advanced tips: put critical facts at the start or end of a long prompt — a Stanford-led study found models use information best in those positions and worst when it is buried in the middle. Pay for a second model to check the work only when a mistake is expensive, such as a contract clause or a payment figure. On OpenAI's API, turn on a hard spend limit for each project, since its spend alerts only notify and don't stop traffic.

Start with one test

Make the cheapest model that passes the five-task test from tip 1 the default for a week, and switch up only when it fails. Then compare the usage page with the previous week's. On the invoice job, routing alone cut the bill from $30 to $3.90 — even after paying Opus 5.5 to redo 300 failures.

Cheaper tokens won't do the job alone, though. The price of a single token has dropped more than 90% since 2023, according to Silicon Data's Token Expenditure Index, yet spending on large language models has doubled since late 2025. Economists call that the Jevons paradox: a resource becomes so cheap that people end up spending more on it in total. A Bain \u0026 Co. brief from June found the same pattern, with token costs halving between December 2024 and 2025 while the number of tokens consumed grew 450%.

The upside is that the same falling prices work for anyone who stops burning tokens for sport. Paired with the habits above, AI gets cheaper and better at the same time.