AI Agents for Lawyers

Glossary

Context Window — Definition for Lawyers

A context window is the text an AI model can hold at once. As of September 2026 it is 1 million tokens on Claude, GPT-6 and Gemini, about 555,000 words.

By Dr. Niklas SchmidtUpdated 13 November 20264 min read
On this page
  1. How big the windows are in September 2026
  2. What a million tokens does not solve
  3. How to work within it

A context window is the maximum amount of text, measured in tokens, that a language model can take in and reason over in a single exchange: your prompt, the documents you attach, the conversation so far and the answer it writes. As of September 2026 the frontier models from Anthropic, OpenAI and Google all offer about 1 million tokens, which Anthropic puts at roughly 555,000 words.

How big the windows are in September 2026

A token is a fragment of text, roughly three-quarters of an English word, and the window is counted in tokens rather than pages; the guide to how large language models work explains what the model does with them. The figures below are the vendors' own, as of 18 September 2026.

ModelContext windowMaximum outputNote
Claude Fable 5.1, Opus 5, Sonnet 51M tokens128KAbout 555,000 words on Anthropic's tokeniser
Claude Haiku 4.5200K64KThe cheap tier at $1/$5 per million tokens
GPT-6 AstraAbout 1.05M128KOff by default in ChatGPT Enterprise until an admin enables it
Gemini 3.1 Pro1M input64K (Gemini 3 Pro figure)API price rises from $2/$12 to $4/$18 above 200K tokens
Grok 4.6500KNot statedReleased 12 August 2026
Muse Spark 1.3, DeepSeek V41MNot statedOpen-weight or partly open alternatives

What this means in practice: a 300-page share purchase agreement with every schedule, or the complete pleadings and exhibits in a mid-sized dispute, fits in one prompt. A data room does not. A diligence exercise of 10,000 documents runs to tens of millions of tokens, which is why Thomson Reuters' Tabular Analysis (10,000 documents against 100 questions) and Legora's Tabular Review (up to 100,000 documents) do not load the corpus at all. They search it and pass the model only the passages that match, the technique described under retrieval-augmented generation.

What a million tokens does not solve

Attention is uneven. Long-context research has shown for years that models recall material at the beginning and end of a long prompt more reliably than material buried in the middle, an effect usually called "lost in the middle". Whether the current generation has fixed it is a claim to test on your own documents. A clause on page 140 of a 300-page bundle deserves a direct question, not an expectation that the model noticed it.

Cost scales with what you load. At Opus 5's $5 per million input tokens, a full window costs about $5 each time the model re-reads it; at Fable 5.1's $10 it is $10, although the 75% cut in cache-read pricing on 1 September 2026 (from $1 to $0.25 per million) makes repeated turns over the same bundle far cheaper. In a chat subscription you never see the invoice; you see the rate limit instead.

The window is not a memory. Context lives for one conversation. GC AI listed "context resets per session" among three gaps in Anthropic's legal plugin (June 2026), and Harvey II (18 August 2026) sells Memory and matter-aware Spaces precisely because a window forgets. The workaround in Claude is a Project, whose uploaded files are re-supplied every session; the Claude Projects setup guide shows how to structure one for a matter.

More context means more to verify. A model that has read 500 pages can misstate any one of them, so the answer needs pinpoint references you can open.

How to work within it

  1. Load the matter, not the firm. Quinn Emanuel's Christopher Kercher told Fortune the breakthrough was "treating Claude like a member of the case team" and feeding it the chronological materials for the case (May 2026). One matter per conversation keeps the window relevant and the confidentiality analysis simple.
  2. State the task at the start and repeat the question at the end. Both positions are where recall is strongest.
  3. Demand clause and page references. "Quote the clause and give the page number" turns a long window into a checkable index; an answer without a pinpoint is an answer you cannot verify.
  4. Use a closed-universe tool when the bundle exceeds the window. Gemini Notebook takes 50 sources on the free tier and 300 on paid plans, cannot browse the open web, and cites passages rather than paraphrasing from memory.
  5. Count before you paste. Non-English text costs more tokens per word, so a German or Czech bundle fills a window faster than its page count suggests.

For due diligence, the practical rule is to decide before starting whether the corpus fits the window or needs retrieval, because the verification workflow differs. Module 3 of the AI Lab for Lawyers course covers file handling and what to load into a window for a live matter.

Frequently asked questions

How many pages fit in a 1 million token context window?

About 555,000 words on Anthropic's tokeniser, which is on the order of 1,500 pages of dense contract text or several complete trial bundles. That is the ceiling, not the recommendation. Load only the documents relevant to the question, because the model re-reads every token on every turn, recall is weakest in the middle of very long prompts, and every page loaded is a page you must be able to check.

Is a bigger context window better than RAG for legal documents?

For one matter's bundle, loading the documents directly is simpler and usually more reliable, because the model sees every page rather than a retriever's selection. For anything larger than the window, such as a data room or a firm's precedent bank, retrieval-augmented generation is the only option: it searches the corpus and hands the model the passages that match. Most legal tools combine the two.

Does the context window include the model's answer?

Yes. Input and output share the window, and vendors cap the output separately: Claude Fable 5.1, Opus 5 and GPT-6 Astra allow 128,000 output tokens, Gemini 3 Pro 64,000. When a long drafting task stops mid-sentence, the output cap is usually the reason. Asking the model to continue in a new turn costs a fresh read of the whole context.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion