New to this? Read it in simple words
- Anthropic made a one-million-token context window available for Claude Opus 4.6 and Sonnet 4.6. It can hold large code projects or many documents.
- In 2023, Anthropic said 100,000 tokens could hold about 75,000 words.
- The real benefit is that users do not repeat the same background in every chat.
- Being inside the window does not mean every fact gets equal attention. The model may still miss a detail.
- Context window
- The amount of text an AI model can read and use at one time.
- Token
- A small piece of text, often part of a word. AI use is priced per token.
A million-token window can hold a great deal of text. It still does not give an AI good judgment or useful long-term memory by itself.
A larger window holds more information
Anthropic made a one-million-token context window available for Claude Opus 4.6 and Sonnet 4.6. Tokens are small pieces of text. This window can hold large code projects, many documents, or a long project history.
In 2023, Anthropic said 100,000 tokens could hold about 75,000 words. Bigger windows allow new uses. They also raise a harder question: what should the system remember, find, shorten, or forget?
Being inside the window does not mean every fact gets equal attention. The model may miss a detail or handle two different instructions badly. Teams still need tests that match their real documents and tasks.
Search tools still have value. They can find the most useful files, show sources, protect private documents, and reduce the amount of text sent with each request. A large window gives more room but does not choose the right information.
Active information, searched project material, and saved facts have different jobs and controls.
Memory must be visible and controllable
A good human partner does not only remember words. They remember goals and decisions. They also know the difference between a lasting fact and a temporary detail.
Long context can cost more and take longer. Old or unrelated information may also distract the model. A good interface lets people control what is active.
A project could show active files, saved facts, and the date each fact was last checked. Users should be able to remove an old preference or keep an important decision with one clear action.
For privacy, forgetting is as important as remembering. Information needs clear borders between projects, people, and companies. Users also need a dependable way to delete it or set an end date.
Sources1
Three useful memory layers
A useful assistant should help people understand what it knows and why it used a certain fact. It should also make corrections easy.
Sources should stay connected to each answer. If advice comes from a meeting note, code file, or earlier decision, the user should be able to open that item and check it.
A clear design may use three layers: active information for the current task, search across a larger archive, and a small list of lasting facts. Users should be able to control these layers without learning how tokens work.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · March 20261M context is now generally available for Opus 4.6 and Sonnet 4.6 Anthropic
- 2
The ideas about interface and memory design are our editorial opinion.
Anthropic made a one-million-token context window available for Claude, letting assistants hold whole projects, though it adds no judgment by itself. How we rate the mood