Context window: how much code fits in 256K tokens

What goes into the context

The context holds system instructions, the whole conversation, attached files and the model answer itself. If the total exceeds the window, old messages get cut or the request fails.

Rough scale

A rough estimate: it depends on language, code style and the model’s tokenizer.

Using a large context well

• Send related files together so the model sees how they interact
• Put what matters at the start and the end; models pay more attention to the edges
• Skip generated and third-party code unless needed
• For very large repositories use an agent that picks files itself

Large context and price

With per-token billing a large context means a large bill, since every request is paid in full. With a Tenzo Cloud subscription a long context costs nothing extra, so you do not have to skimp on what the model should see.

Example: a pull request review

A typical PR touches 5–15 files. To understand it, the model needs not only the diff but neighbouring code: interfaces, call sites, tests. That is easily 20–40K tokens. A 256K window fits the PR plus most of the affected module, so the model sees the whole picture instead of pieces.

When a large context helps most

• Exploring an unfamiliar codebase
• Comparing versions of long documents
• Analysing hours of incident logs
• Working with large spreadsheets and exports