English

A million tokens of context, and what that actually is

Anthropic released a model yesterday whose context window is five times larger than before. This is the most practically useful change in months, for one narrow reason.

On 5 February 2026 Anthropic released Claude Opus 4.6, with a context window of one million tokens, up from two hundred thousand, and reasoning that engages only when the task calls for it.

Context window is jargon, and it is worth translating, because this particular number matters more to a small company than most model specifications do.

What a context window is

How much you can put in front of the model in one go, and how much of the conversation it can still see. Not what it knows — what it can look at right now.

A million tokens is roughly seven hundred thousand words. Your entire general terms and conditions, a full technical manual, three years of correspondence with one customer, or the complete documentation of a machine.

Why this is more useful than it sounds

Because it removes the most common failure from the 29 September post: the model gave a wrong answer because the right answer was in a document it could not see.

Until now, working with large documents meant chopping them up and hoping the relevant piece was in the part you sent. That was a real source of quiet errors, and it is the kind of problem nobody diagnoses because the output looks confident either way.

What it does not do

It does not mean the model reads a large document well. Attention thins out across very long inputs, and the useful discipline is unchanged: ask a specific question, not « summarise this ».

It does not remove the cost question. A large input costs more than a small one on a per-use contract, and a habit of pasting an entire manual for every question will show on the bill. The 15 September budget advice applies.

And it is not a substitute for connecting the model to your documents properly. Pasting is a manual act; a system that can find the right document is a different thing.

Worth testing if

You have a long document that people repeatedly need answers from — a contract, a standard, a manual, a regulation. That is a concrete task, it is on your twenty-example list or should be, and next week’s post works through it.

← Back to all posts