Author: mollygoldreich

If you’ve ever opened an API bill and thought “wait, why is this so much higher than I expected,” you’re in good company. AI pricing looks simple on the surface, but it hides a few quirks that catch almost everyone off guard the first time.

Tokens, not words

Here’s the first thing that trips people up. AI providers don’t charge per word. They charge per token, and a token is roughly four characters, or about three quarters of a word in English. A word like “hamburger” might get split into two or three separate tokens depending on the model, while short common words like “the” usually stay whole.

Why does this matter? Because it means your bill isn’t tied to something intuitive like word count. It’s tied to how a specific model happens to slice up your text, and that slicing works differently from provider to provider. We go deeper into exactly how that splitting works in How AI Models Actually Read Your Text, if you want the full picture.

Input costs less than output, usually by a lot

Nearly every provider charges more for what the model generates than for what you send it. Sometimes it’s double, sometimes five times as much. This is part of why a chatbot that writes long, detailed responses racks up costs faster than one you’re mostly using to summarize or extract short answers from documents you provide.

If you’re building something where the model talks a lot, keep an eye on this ratio. It usually matters more than people expect going in.

Not all models are priced the same, even from one company

Most providers give you a lineup to choose from. OpenAI has its flagship GPT models alongside smaller, faster mini versions. Anthropic has Claude Opus at the top end and Haiku for speed and volume. Google does something similar with Gemini Pro and Flash.

The tempting move is to reach for the biggest, smartest model by default. But a lot of everyday tasks, classification, tagging, short rewrites, don’t actually need that much horsepower. Picking the right tier for the job is honestly one of the easiest ways to cut your bill without losing quality where it counts. We cover more cost-cutting habits like this in 10 Ways to Actually Cut Your AI API Costs.

Why the same prompt costs different amounts on different platforms

Two things stack together here. First, each provider just sets its own price per token. Second, and this part surprises people, each provider’s tokenizer counts your text differently in the first place. So even before you get to pricing, the raw token count for the exact same sentence can vary between GPT and Claude and Gemini. For a full side-by-side breakdown, check out GPT vs Claude vs Gemini: What Actually Costs More.

This is exactly the problem we built How Many Tokens to solve. Paste your text into our calculator, and you’ll see the token count and estimated cost across 45+ models side by side, so you’re comparing real numbers instead of guessing.

The takeaway

Once you understand tokens, input versus output pricing, and how tokenizers differ, AI costs stop feeling random. Check your numbers before you scale a feature, and you’ll avoid most of the surprises that catch people off guard. If you’ve still got questions after all this, our FAQ page covers a lot of the common ones.

Picking between GPT, Claude, and Gemini usually comes down to more than which one seems smartest. Cost per token, context window size, and how the pricing is structured all shape which one actually makes sense for what you’re building. Here’s how they stack up.

A quick rundown of the three

OpenAI’s GPT models are the most widely adopted option out there, with a range running from lightweight mini models up to flagship reasoning models. Pricing scales pretty predictably with capability, and the tooling and documentation around it are some of the most mature in the space.

Anthropic’s Claude models lean into strong reasoning and handling longer context well. The lineup runs from the fast and affordable Haiku models up through Sonnet and Opus for heavier lifting. Claude has built a solid reputation for careful, well thought out output on complicated tasks.

Google’s Gemini models plug directly into Google’s broader ecosystem and often lead the pack on sheer context window size, with some versions handling enormous inputs. Like the others, Gemini’s lineup ranges from fast and cheap up to top tier reasoning.

Context window matters just as much as price

A cheaper price per token doesn’t automatically mean a cheaper bill. If a model has a smaller context window, you might end up making more calls or truncating content just to fit, which adds up fast. Sometimes a provider with a much larger context window actually works out cheaper overall for document heavy work, even at a higher per token rate, simply because you need fewer round trips to get the job done. For more on how context windows actually work under the hood, see How AI Models Actually Read Your Text.

The input and output split isn’t the same everywhere

All three charge less for input than output, but by how much varies. Some price output at roughly double the input rate. Others charge significantly more. If you’re building something that generates long responses, like content writing or code generation, this ratio ends up mattering more than the headline input price most people compare first. Our pricing guide walks through this split in more detail.

Matching the provider to what you’re actually doing

• High volume, simple tasks like classification or tagging: the lightweight tier from any of the three tends to land in a similar price range, so it’s worth checking your actual numbers rather than assuming.
• Long documents or big codebases: a bigger context window can mean fewer total calls, which sometimes offsets a higher per token price.
• Complex reasoning or careful writing: the flagship tiers across all three are priced fairly close together for comparable quality, so output quality becomes the real deciding factor over cost.

Once you’ve picked a direction, 10 Ways to Actually Cut Your AI API Costs has practical steps for keeping whichever provider you choose as affordable as possible.

Skip the guessing

Marketing pages rarely make it easy to compare real costs side by side. The most reliable approach is to take an actual sample of your own prompts and run the numbers across providers directly.

Paste your text into our free tool and see the token count and estimated cost across GPT, Claude, Gemini, and 40 plus other models in one place, so you’re comparing apples to apples before picking a provider. Curious who’s behind this project? Read more about us here.

Type a message to an AI model and it doesn’t see words the way you do. It sees tokens, small chunks of text turned into numbers it can process. Once you understand this, a lot of the confusing parts of AI pricing and limits suddenly make sense.

Breaking down a sentence

Take something simple: “The quick brown fox jumps.” You’d read that as six words. A tokenizer might see something closer to:

The | quick | brown | fox | jump | s | .

Notice “jumps” got split in two. That’s because tokenizers are built from patterns found in huge amounts of text, not from a dictionary of whole words. Common words often stay whole. Less common ones, or words built from smaller parts, get chopped up.

Why the same sentence counts differently depending on the model

Every provider trains its own tokenizer on its own data. That means the exact same sentence can come out to a different token count depending on whether you’re using GPT, Claude, or Gemini. A prompt that’s 100 tokens on one might be 115 on another, purely because of how each one happens to split the text.

This is one of the most overlooked reasons that comparing prices across providers can be misleading if you’re only looking at the sticker price per token. Two providers can charge the same rate and still cost you differently, just because one counts more tokens for your text than the other. Our pricing guide and provider comparison both dig into how this plays out with real numbers.

Non-English text usually costs more

Most tokenizers are trained on data where English dominates. That means text in other languages, especially ones with different scripts like Chinese, Japanese, or Arabic, often needs more tokens to say the same thing. If you’re building for a non-English audience, this is worth factoring into your cost estimates early.

The hidden tokens you never see

Beyond the actual words in your prompt, models also use invisible structural tokens behind the scenes, marking where a conversation turn starts and ends, separating instructions from user input, that kind of thing. It’s a small overhead, but it’s there on every single call whether you notice it or not.

Why any of this matters day to day

Once tokenization clicks, a few things stop being mysterious:

• Why your token count is always higher than your word count
• Why the same document costs differently depending on which provider you use
• Why a “128K context window” doesn’t map neatly onto some number of pages
• Why keeping prompts short and direct actually saves real money, something we cover in more detail in 10 Ways to Actually Cut Your AI API Costs

See it happen in real time

The easiest way to actually get a feel for this is to watch it happen. Paste any text into our free calculator and see exactly how it breaks apart across 45+ models. No signup needed. For quick answers to related questions, our FAQ page has more detail on context windows, pricing, and how the calculator works.

Most people don’t overspend on AI because the technology is expensive. They overspend because of a few habits that quietly add up. Fix these and you’ll usually see a real difference on your next invoice.

1. Stop defaulting to the biggest model

It’s tempting to just use the most powerful option every time. But plenty of tasks, sorting, tagging, short summaries, run perfectly well on a smaller, cheaper model. Save the flagship models for the jobs that actually need deep reasoning. Our token pricing guide breaks down how much that choice actually saves across providers.

2. Trim the fat out of your prompts

Long system prompts cost you tokens on every single call, forever, whether they’re helping or not. Go back through yours every so often and cut anything that isn’t clearly earning its keep.

3. Look into prompt caching

If you’re sending the same background context over and over, check whether your provider supports caching for repeated content. It can cut costs noticeably once you’re making a lot of similar calls.

4. Set real limits on output length

An open-ended max token setting invites longer, pricier responses than you probably need. Cap it based on what the task realistically requires. Since output tokens are usually priced higher than input, this is one of the highest-leverage changes you can make.

5. Batch what doesn’t need to be instant

Some providers offer cheaper batch processing for anything that doesn’t need a real-time response. If part of your workflow can wait a bit, this is free money left on the table if you’re not using it.

6. Summarize before the expensive step

Feeding huge documents straight into your best model gets pricey fast. Run a cheap summarization pass first, then send the condensed version to the model doing the harder reasoning work.

7. Watch your conversation history

In multi-turn chats, previous messages often get resent as context every single time. Left unchecked, this quietly balloons your token count the longer a conversation runs. Trim or summarize history periodically. If you want to understand exactly why this happens, How AI Models Actually Read Your Text explains how context gets counted under the hood.

8. Test before you scale

Before rolling something out to thousands of users, paste a realistic sample of your actual prompts into a token calculator. It takes two minutes and tells you what you’re actually about to pay at volume, instead of finding out from the invoice.

9. Break down usage by feature

Look at your usage per feature, not just the total. It’s surprisingly common to find one small, underused feature quietly eating a huge share of your budget.

10. Revisit your model choices every few months

Pricing and capability shift often. The model that made sense six months ago might have a cheaper option now that does just as good a job. See how the major providers currently stack up in GPT vs Claude vs Gemini: What Actually Costs More.

Start with visibility

None of this works if you don’t know your numbers to begin with. That’s the whole reason our calculator exists: paste your text, see the token count and cost across 45+ models instantly, and make the call before you build, not after. You can read more about why we built this tool if you’re curious about the reasoning behind it.

Welcome to WordPress. This is your first post. Edit or delete it, then start writing!