Gains:
- Ability to explain the concepts of token, context window and multimodality in everyday language
- Diagnose whether a task requires broad context or multimodality
- Ability to take into account how these concepts affect cost and model selection
So far we are familiar with the providers and model types. However, for the right model selection, it is also necessary to understand how the models work "inside"; because price, capacity, and availability decisions all rely on three core concepts: token, context window, and multimodality. Someone who does not know these concepts cannot read the price sheet and wonder "will this document fit the model?" cannot answer the question and cannot see when visual processing is essential. By the end of this unit, you will be able to explain these three concepts in everyday language, diagnose whether a job requires broad context or multimodality, and take into account their impact on cost.
Token: Unit Used by the Model Instead of Word
Artificial intelligence models process text not word by word, but in small pieces called tokens. A token; It can be a word, part of a word, a punctuation mark, or a space. To make a rough calculation: In English, an average token is about four characters, and 100 words are about 130-150 tokens. In Turkish, this rate may be slightly higher due to suffixed and long words; In other words, a Turkish text with the same meaning consumes slightly more tokens than English.
Why is this important to know? Because everything is charged and measured in tokens. The text (input) you send to the model and the response (output) produced by the model are counted as separate tokens. Both your invoice and the capacity of the model are in tokens.
Tip: To get a rough estimate of how many tokens a text has, divide the number of characters by four. A 4,000 character email is approximately 1,000 tokens. In Turkish, assume a little higher to stay on the safe side.
Context Window: Model's "Short Term Memory"
The context window is the total amount of tokens that the model can keep in mind at a time. Input and output must fit together in this window. Think of it like the amount of paper you can spread on a table at once: The paper that doesn't fit on the table is out of your field of vision at that moment.
If a model's context window is 200,000 tokens, this equates to approximately 150,000 words, or several medium-sized books. 1 million tokens can process much larger documents as a whole. Some powerful models today (e.g. Claude Opus 4.8 and Sonnet 5) offer 1 million token contexts, while lightweight models often come with smaller windows (e.g. 200,000 tokens for Haiku 4.5).
Why is it important? Because if a document does not fit in the context window, the model cannot see it all together. You can't give a 500-page contract in its entirety to a 200,000 token model; You either divide the document into parts (which may lose the relationship between the parts) or choose a model with a broader context.
Caution: It is not wise to fill in every document as it is just because the context window is large. The more tokens you put into the window, the more you pay, and sometimes the model can miss the middle details in very long texts. It is often cheaper and more accurate to send only the part that works.
Context Window is a Decision Criteria
Quest
Approximate required context
Appropriate level
Short email response
several hundred tokens
lightweight
One page summary
several thousand tokens
Light/balanced
40-page report analysis
~30,000 tokens
Balanced/strong
300-page contract review
~250,000 tokens
Powerful with broad context
Process hundreds of documents at once
500,000+ tokens
Broadest context
This table answers the question "which model?" It shows that part of the question is answered directly by document size. It is a waste to buy an expensive model with a large context for short jobs; A model with a small window is inadequate for large documents.
Multimodality: Going Beyond the Text
Multimodality is the ability of a model to handle multiple types of data (image, audio, sometimes video), not just text. "Modal" here means "data type"; multimodal means "many kinds".
- Text-only model: Reads and writes only written text.
- Multimodal model: Can read a photo of a bill, interpret a trend on a graph, decode a handwritten note, sometimes transcribe a voice recording.
Why is it important? Because the nature of your business may require multimodality. An accounting team that extracts amounts from invoice images, an e-commerce team that classifies product photos, or an insurance team that evaluates damage photos cannot do this job with a model that only reads text. Visual processing is essential to these tasks.
Three Realistic Cases
Case 1 — Token cost surprise. A content team also sends the model a 20-page “brand guide” document as they produce each blog post. This guide is about 15,000 tokens. They produce 2,000 articles a month; So just sending the guide over and over again means 30 million tokens of input. By shortening the guide and sending only the relevant section (about 2,000 tokens), they reduce the input to one-seventh and significantly reduce the bill.
Case 2 — Context window is not enough. A law firm wants to have a 280-page merger agreement reviewed. The first model they tried had a binding of 200,000 tokens and the document did not fit; When the document is torn apart, the model cannot notice that an article in the first section contradicts an article in the last section. When it switches to a model with 1 million token context, it reads the document as a whole and catches the contradiction. Here the context window directly determined accuracy.
Case 3 — Multimodality is a must. An insurance company wants to make a preliminary assessment from vehicle damage photographs. This is impossible with a model that only reads text; A multimodal model that "sees" the photograph is required. When they switch to a multimodal model, they establish a pre-screening system that roughly classifies the location and severity of damage in the photo and directs the expert. They note, however, that a human expert makes the final decision; because the model is a preliminary screening tool, not the final word.
Weak Prompt / Strong Prompt
Weak prompt:
Summarize this document. [too long document pasted]
Powerful prompt:
Summarize the 40-page report below. First estimate the approximate number of tokens in the report and tell us whether it fits within the context window of a typical balanced model. Then give the summary in this format: 5-item executive summary + 3 risky findings. Use only the information in the report; Mark the part you are not sure of as "not clear in the report". [report]
The powerful prompt is both token/context aware and controls the format and verifiability of the output.
Copiable Templates
1. Token prediction:
Estimate the approximate number of tokens for the following text and interpret this in terms of input cost: "[TEXT]". Round it up a little, taking into account that it is Turkish.
2. Context availability check:
My job is to process the following documents: average [PAGES] of [QTY] documents per month. What context window does this size require? Which of the light/balanced/strong levels would be sufficient? What risk arises if it needs to be dismantled?
3. Multimodality diagnosis:
My workflow: [DESCRIPTION]. Is there visual, audio or video processing in this stream? If so, is a multimodal model required, or can it be converted to text (OCR/transcription) and solved with the text model? Compare the pros/cons of the two paths.
4. Context optimization:
I send a document to the model with every request, but most of it is unnecessary. How do I extract the parts that are actually required for [TASK] from this document? Suggest 3 concrete ways to reduce input and indicate estimated token savings.
Common mistakes
- Mistaking token for a word: Token is different from a word; Invoice and capacity are always in tokens, calculating in words is misleading.
- Thinking the context window is infinite: Every model has a limit; If the document does not fit, the model cannot see the whole.
- Using unnecessary large context: Buying an expensive model with a large context for a short job; or fill out huge documents for every request and pay in vain.
- Assuming multimodality: Not every model can process visuals; For visual work, start without confirming that the model is multimodal.
- Forgetting the token load of Turkish: Turkish texts generally consume slightly more tokens for the same meaning; ignoring this in cost estimation.
In summary
- A token is the small unit in which the model processes text; input and output are measured and invoiced separately as tokens.
- The context window is the total tokens the model can keep in mind at a time; document size directly affects model selection.
- Multimodality is the ability of the model to process non-text data such as visual/audio; It is essential for visual works.
- These three concepts are both relevant to the question "which model?" as well as “how much does it cost?” It forms the basis of the questions.
Application task
Choose a recurring AI task in your business. Have the 1st template (token prediction) calculate how many tokens a typical input is in this task. Then determine which level is sufficient with template 2 (context availability check). If you have a visual job, clarify whether a multimodal model is required with the 3rd template (multimodality diagnosis). We will use the resulting token estimate in the cost calculation of the next unit.
checklist
- [ ] I know the concept of token and how to roughly estimate how many tokens a text has.
- [ ] I can explain the context window with a "table/memory" analogy.
- [ ] I can evaluate whether a document will fit into a model.
- [ ] I can diagnose when multimodality is essential.
- [ ] I take into account the token overhead of Turkish and the cost of unnecessary context.