The Claude context window — the total amount of text a model can hold in a single conversation — was 200K tokens for so long that “200K” became shorthand for Claude itself. That era is over: as of 2026, every current Opus, Sonnet, and Fable model carries a 1 million token context window, with only Haiku 4.5 remaining at the classic 200K. This article explains what the context window actually is, what changed between the 200K era and today, what a million tokens really holds, and how to get good results out of very long contexts.
TL;DR:
• The context window is Claude’s working memory: everything in the conversation — your messages, attachments, and Claude’s replies — must fit inside it.
• 200K tokens (~150,000 words) was the long-time Claude standard; current Opus/Sonnet/Fable models are 1M tokens, while Haiku 4.5 stays at 200K.
• 1M tokens ≈ 750,000 words: an entire mid-size codebase, several novels, or years of meeting notes in one conversation.
• Context is read capacity, not write capacity — max output is 128K tokens (64K on Haiku).
• Long contexts work best with good structure: documents first, questions last, and fresh chats when topics change.
What Is a Context Window?
Think of the context window as Claude’s desk space. Everything Claude can “see” while answering — the system prompt, every message you’ve sent, every file you’ve attached, and every reply it has already given — sits on that desk, measured in tokens. One token is roughly 4 characters of English text, so 1,000 tokens is about 750 words. (If tokens are new to you, our explainer on Claude AI token limits covers the unit in depth.)
Two properties follow from this design. First, the window is a hard ceiling per conversation: once a chat’s accumulated history approaches it, something has to give — on claude.ai you’re prompted to start a new conversation. Second, the window is shared between reading and writing: input and output both occupy the same space, though output has its own separate per-response cap.
The 200K Era — Why Everyone Says “200K”
If you search for “claude context window,” half the results still say 200K — and for good reason. A 200,000-token window was the Claude standard for a long stretch of the model line’s history, spanning multiple generations of Opus, Sonnet, and Haiku. At the time it was a headline advantage: 200K tokens is roughly 150,000 words, or about 500 pages — enough for a full technical book, a legal contract stack, or a substantial slice of a codebase, back when much of the competition offered a fraction of that.
So when this article’s title says “200K Explained,” here’s the honest 2026 answer: 200K is now the floor, not the ceiling. It survives on exactly one current model — Claude Haiku 4.5, the speed-and-cost tier — while everything above it has moved to 1M. Any guide still quoting 200K as Claude’s overall limit is describing the previous era.
Claude Context Window Sizes in 2026
| Model | Context window | Max output | Approx. capacity in words |
|---|---|---|---|
| Claude Fable 5 | 1M tokens | 128K | ~750,000 words |
| Claude Opus 4.8 | 1M tokens | 128K | ~750,000 words |
| Claude Opus 4.7 | 1M tokens | 128K | ~750,000 words |
| Claude Sonnet 5 | 1M tokens | 128K | ~750,000 words |
| Claude Sonnet 4.6 | 1M tokens | 128K | ~750,000 words |
| Claude Haiku 4.5 | 200K tokens | 64K | ~150,000 words |
The same windows apply whether you reach the models through claude.ai or the API — though on claude.ai your plan’s usage budget usually becomes the practical constraint before the window does, a distinction we unpack in Claude AI limits explained. For how the models differ beyond context size, see our Claude models comparison.
What Actually Fits in 1 Million Tokens
A million tokens is around 750,000 words of English prose. Concretely, that’s:
- An entire mid-size codebase. Tens of thousands of lines across hundreds of files, loaded in one shot — the use case that changed how I work with Claude Code, since the model can hold the whole architecture in view instead of a curated slice.
- Several full novels. War and Peace is roughly 587,000 words; a 1M window fits it with room for a sequel and your questions.
- A complete document corpus. A year of board minutes, a full contract data room, or dozens of research papers, cross-referenced in a single conversation.
- Very long agentic sessions. Multi-hour Claude Code or Research runs where the trail of tool calls, file reads, and intermediate results no longer forces aggressive summarization.
The one thing a 1M window does not give you is a 1M-token answer: output remains capped at 128K tokens per response (64K on Haiku). The window is about how much Claude can read and reason over, not how much it writes back.
How to Use a Long Context Window Well
After extensive long-context work, these habits consistently produce better results:
- Documents first, questions last. Put the bulk material at the top of your prompt and the instruction at the bottom. Claude handles “here are 400 pages… now answer this” better than the reverse.
- Ask for citations. Have Claude quote the specific passage or file it’s drawing from. This grounds answers in the actual context and makes hallucinations easy to spot.
- Structure the input. Label documents clearly (headers, filenames, XML-style tags). A million tokens of unlabeled soup is much harder to reference precisely than the same material with structure.
- Don’t hoard context you don’t need. Relevant-only context beats everything-just-in-case. More material means more places for attention to land on the wrong thing — and on the API, more input cost.
- Use prompt caching for repeated context on the API. If the same big corpus backs many requests, caching drops the effective input price to roughly 0.1× on cache reads. The economics are covered in our Claude API pricing guide.
- Start fresh when the topic changes. A huge window makes it tempting to run one eternal thread. Resist — carryover from stale topics pollutes answers and burns your usage budget.
Long-Context Caveats — Where Quality Degrades
A 1M window doesn’t mean uniform performance across all million tokens, and it’s worth being honest about the failure modes:
- Mid-context recall is the weak spot. Details buried in the middle of a very long context are recalled less reliably than material near the beginning or end. Current models are far better at this than the 200K-era ones, but the pattern hasn’t vanished — put critical facts where they’re easy to find, and ask for quotes.
- Aggregation is harder than retrieval. “Find the clause about termination” works brilliantly at any length. “Summarize every obligation across all 40 contracts” is where long-context answers get lossy — break aggregation tasks into passes.
- Latency and cost scale with input. Processing hundreds of thousands of tokens takes noticeably longer than a short prompt, and on the API you pay for every input token each time (unless cached).
- The window is not memory. Context vanishes when the conversation ends. Persistent recall across chats is what claude.ai’s Memory and Projects features are for — a different mechanism entirely.
What People Get Wrong About the Context Window
- “Claude’s context window is 200K.” Outdated for everything except Haiku 4.5. Current Opus, Sonnet, and Fable models are all 1M.
- “Tokens are words.” A token is ~4 characters; 1M tokens is ~750,000 words, and code tokenizes less efficiently than prose.
- “A bigger window means longer answers.” Output is capped separately at 128K tokens regardless of window size.
- “The context window is my usage limit.” The window bounds one conversation’s size; your plan’s session and weekly budgets bound how much you can send over time. Different ceilings, hit for different reasons.
- “Filling the window is free on claude.ai.” Every token of history is re-processed each turn and drawn from your usage budget — big contexts spend it fast.
Anthropic documents current model specs, including context and output limits, in its official models overview.
FAQ
What is Claude’s context window in 2026?
Current Claude Opus (4.7 and 4.8), Sonnet (4.6 and 5), and Fable 5 models all have a 1 million token context window. Claude Haiku 4.5 keeps the long-time 200K standard. A million tokens is roughly 750,000 words of English text.
Is Claude’s context window still 200K?
Only on Haiku 4.5. The 200K window was the Claude standard for a long time across multiple model generations, which is why so many guides still quote it, but every current Opus, Sonnet, and Fable model now offers 1M tokens.
How many words fit in a 1M token context window?
About 750,000 words of English prose, using the rule of thumb that one token is roughly 4 characters. That is enough for an entire mid-size codebase, several full novels, or a year’s worth of company documents in one conversation.
Does a bigger context window mean longer responses?
No. The context window is read capacity – how much Claude can hold and reason over. Output is capped separately at 128K tokens per response on 1M-context models and 64K on Haiku 4.5.
Does Claude remember things outside the context window?
Not within a plain conversation – once a chat ends or content falls outside the window, it is gone. Persistent recall across conversations comes from claude.ai features like Memory and Projects, which work differently from the context window itself.
Does performance degrade with very long contexts?
Somewhat. Retrieval of specific facts stays strong, but details buried mid-context are recalled less reliably than material near the start or end, and broad aggregation tasks across huge inputs can get lossy. Structuring documents clearly and asking Claude to quote its sources mitigates most of it.
Next Steps
The Claude context window went from an impressive 200K to a routine 1M, and the winning skill shifted with it: not squeezing material in, but structuring what you load so Claude can use it precisely. To see how the window interacts with what you’re actually charged for, read our breakdown of Claude AI token limits next — and find more quick answers in the FAQ hub.