Everyone's talking about token optimisation. Rightly so, with the likes of Fable getting so damn expensive.

Cutting consumption / staying under the limit is half of every AI conversation lately, and the word “token” comes up in these conversations.

Unless you’re an AI product builder, you probably don’t truly understand what it means, and you can't optimise something you don't understand.

But I think it’s high time to talk about this.

The bills have turned brutal: Uber burned through its entire 2026 AI budget in 4 months, and Meta started rationing tokens across 6,000 staff.

You're on the same meter they are, just with smaller numbers.

So before any of the tricks, I figured we can start with the basics.

But before that, some catchup:

Partner with us

1. GPT for Work Tokenizer: Paste any text and watch it split into tokens, with live counts across GPT, Claude, Gemini and Grok. It's a fast way to see how many tokens your prompts actually use.

2. OpenAI Tokenizer: OpenAI's official tool that shows how your words get chopped into tokens and how the count adds up. Type a sentence, delete a word, and watch the number move.

Think of it like a taxi meter.

The moment you send a message the meter starts, and it keeps running on every word you send and every word the AI sends back.

The longer the ride, the bigger the fare.

A token is roughly a piece of a word.

Short, common words like "the" or "is" are one token each, and longer ones get split into chunks, so "unforgettable" becomes "un", "forget" and "table".

As a rule of thumb, 75 words of English is about 100 tokens.

You type words, but the AI never sees them.

Before it does anything, your text gets chopped into tokens, and each token becomes a number, because numbers are all the model works with.

When it replies, it produces numbers too, which get turned back into the words on your screen.

That chopping step is called tokenisation, and every model does it a little differently, which is why the same sentence can cost a different number of tokens on Claude than it does on ChatGPT or Gemini.

Not every token does the same job or costs the same.

The split that matters most is input versus output.

So the bill comes from long answers, even when the question that triggered one was a single line.

A few quieter ones add up too.

Cached tokens are a discount most people never claim: send the same big chunk of context every time and modern models can store and reuse it instead of re-reading it.

Reasoning tokens are the hidden thinking some newer models do before they answer, which you pay for even though you never see it, so a 200-word reply can have thousands of invisible tokens sitting behind it.

And anything you upload becomes tokens as well, with images the worst offender, where 1 screenshot can cost more than several pages of text.

In most chat apps the AI has no memory between turns.

To stay coherent it re-reads the entire conversation every single time you hit send, so every message you send pays to reprocess the whole chat on top of whatever you just added.

Once you can see the meter, a few familiar annoyances stop being mysteries.

That drift from sharp early answers to vague, generic ones comes from the context filling up and squeezing out room to think.

Hitting your usage limit way ahead of schedule usually traces back to a file or two you uploaded, each swallowing thousands of tokens.

The slowdown at the tail end of a long chat is every new message hauling the entire history behind it.

And the lift you feel starting fresh is real: a new chat begins at zero, with the whole context window free again.

Lowering your token use comes down to a handful of habits, simplest first.

And if you work with the API or coding agents:

Most token advice is noise.

The ten-point checklists and the calculators are built for people running API bills.

If you're just typing into a chat box, almost none of it is aimed at you.

Two habits carry the load: start a new chat when the subject changes, and ask for a short answer when you don't need a long one.

Caching and structured outputs are real, but they live in code you'll never open, and dressing them up as everyday advice just hands ordinary users one more thing to feel behind on.

You don't need to count tokens. You need those two reflexes, and the quiet sense that the meter is always running.

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

Did you recently see an ad for Roku Ads Manager in a newsletter? We’re running a short brand lift survey to understand what’s actually breaking through (and what’s not).

It takes about 20 seconds, the questions are super easy, and your feedback directly helps us improve how we show up in the newsletters you read and love.

If you’ve got a few moments, we’d really appreciate your insight.

Take the survey.

#AD 2

While many businesses are solving problems at lightspeed, their payroll systems seem to stay stuck in the past. Deel's free Payroll Toolkit shows you what's actually changing in payroll this year, which problems hit first, and how to fix them before they cost you. Because new compliance rules, AI automation, and multi-country remote teams are all colliding at once.

Check out the free Deel Payroll Toolkit today and get a step-by-step roadmap to modernize operations, reduce manual work, and build a payroll strategy that scales with confidence.

Download the Toolkit for free today