# What does Claude's new 100K context window mean, and what can it read at once?

Date: 2023-05-12 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-claude-100k-context-window/

**Answer:** Anthropic has raised the amount of text Claude can take in at once from about 9,000 tokens to 100,000 tokens, roughly 75,000 words, or a whole short novel. That lets a single request hold hundreds of pages, so Claude can summarize, compare and answer questions across long documents in under a minute.

Yesterday, May 11, 2023, Anthropic announced that its assistant Claude can now work with 100,000 tokens of text in one go. Before this change the limit was about 9,000. In everyday terms, Claude has gone from holding a long essay in view to holding an entire book.

Anthropic made the announcement in a short post on its website. The larger window is available through its API, the programming interface that companies use to build Claude into their own products. Businesses that are not yet using Claude can ask for access. The company frames the change for professional work: contracts, financial filings, research papers, technical manuals and whole software projects.

## What is a context window?

A language model like Claude does not read a document the way you do, page by page over several evenings. Everything it considers for a reply must fit into one block of text sent with the request: your instructions, any documents you paste in, and the conversation so far. That block is the **context window**.

The window is measured in **tokens**, the small pieces a model splits text into. A token can be a whole short word or part of a longer one. TechCrunch gives the example of "fantastic," which can be cut into three pieces. Anthropic's own figures show the usual ratio: 100,000 tokens is about 75,000 English words, so one token is roughly three quarters of a word.

Whatever falls outside the window does not exist for the model. That is why chatbots with small windows lose track in long conversations. After a few thousand words, they can forget the first instructions and drift off topic. A bigger window does not give Claude a memory from one day to the next, but it does let much more fit into a single session.

## How big is 100,000 tokens?

Here is a simple comparison, using the three-quarters ratio above. The word counts are rounded estimates, not official limits.

| Window | Approximate words | Roughly enough for |
|---|---|---|
| Claude before this week: 9,000 tokens | about 6,700 | A long article or a short report |
| GPT-4's largest version: 32,000 tokens | about 24,000 | A long report or a thin booklet |
| Claude now: 100,000 tokens | about 75,000 | A short novel, or hundreds of pages of documents |

TechCrunch notes that GPT-4's 32,000-token version held the previous record. Anthropic's new limit is about three times larger.

Anthropic also offers a few other ways to picture the size. An average reader would need more than five hours to get through 100,000 tokens, and far longer to absorb and analyze them. In audio terms, it equals about six hours of speech once transcribed.

## The Great Gatsby test

To show the window in action, Anthropic loaded the full text of F. Scott Fitzgerald's *The Great Gatsby*, about 72,000 tokens, into Claude Instant, the faster and lighter version of Claude. It changed one sentence, so the narrator, Mr. Carraway, became "a software engineer that works on machine learning tooling at Anthropic." Then it asked the model what was different. Claude found the changed line in 22 seconds.

The test is a little playful, but it checks something real. To spot one altered sentence, the model has to treat the whole novel as one piece of text and notice what does not belong.

Anthropic also points to a demonstration by AssemblyAI, a company that turns speech into text. AssemblyAI transcribed a long podcast, about 58,000 words, and then used Claude to summarize the episode and answer questions about it.

## Why it matters this week

Until now, a common way to work with long material was to break it into small chunks, store them, and pull out only the chunks that seem relevant to each question. This method is often called **vector search**: each chunk is turned into a list of numbers, and the system fetches the chunks whose numbers sit closest to the question. It works, but it can miss answers that depend on many parts of a document at once.

Anthropic argues that for complex questions, simply placing all the documents in the prompt will often work better. Its list of suggested uses shows the idea:

- summarizing and explaining dense papers or financial statements;
- weighing a company's risks and opportunities from its annual reports;
- setting out the pros and cons of a proposed law;
- finding themes and arguments across a stack of legal documents;
- answering technical questions from hundreds of pages of developer documentation;
- loading a whole codebase so Claude can build on it or change it.

There are limits. TechCrunch reminds readers that Claude still forgets everything between sessions, and that it gives every piece of text in its window the same weight, the way a reader might give a footnote and a headline equal attention. Some researchers think those problems will need new model designs. And for now, the feature reaches developers and business partners through the API rather than the general public.

> **Try it:** Pick a document you know very well, such as a short story or a work handbook. Change one sentence in a small but meaningful way. Then write down the exact question you would ask an AI to find the change, and decide how you would check whether its answer is right.

## How can you make good use of a very long context window?

Start with a clear task before you paste anything. A long window lets you include a lot, but the model still needs to know what you want: a summary for a manager, a list of risks, or answers to five specific questions. Put the instructions near the beginning, describe the documents briefly, and say what format you want back.

Ask questions that use the whole text. A long window shines when the answer depends on several places at once: "Where do these two contracts disagree?" or "Which chapters mention the budget, and how do the figures change?" For a single fact on one page, a short excerpt is faster and cheaper.

Finally, check the answers. Ask Claude to point to the section or quote it relied on, then look it up yourself. A model that can read 75,000 words in under a minute can still misread one of them, so treat its reply as a fast first draft from a capable assistant, and keep your own judgment for the final decision.

## Sources

- Introducing 100K Context Windows: https://www.anthropic.com/index/100k-context-windows
- Anthropic's latest model can take 'The Great Gatsby' as input: https://techcrunch.com/2023/05/11/anthropics-latest-model-can-take-the-great-gatsby-as-input/
- Introducing Claude: https://www.anthropic.com/news/introducing-claude