Tokens, context windows, and what “running out” means

The short answer

A token is roughly 4 characters of text — about three-quarters of a word. The context window is how much text Claude can hold in its head at once. Claude’s current limit is 200,000 tokens, which is about 500 pages. “Running out” means the conversation is too long for Claude to remember the start.

Tokens, in practical terms

A token is the unit AI models read and write in. It’s not a character and not a word — it’s somewhere in between.

For English text, a useful approximation:
– 1 token ≈ 4 characters
– 1 word ≈ 1.3 tokens
– 100 words ≈ 130 tokens
– 1 page of normal text ≈ 400 tokens

For Indian languages and code, the math shifts — Hindi or Tamil text uses more tokens per word, and code uses fewer. The 4-characters-per-token rule is for plain English prose.

What the context window actually limits

When you’re chatting with Claude, the entire conversation — your messages plus Claude’s responses plus any files you’ve uploaded — counts against the context window.

Claude’s current limit is 200,000 tokens. That’s roughly:
– 500 pages of text
– A 200-page PDF plus a 100-page reference doc plus a long conversation
– About 150,000 words

Most people don’t hit this limit. The ones who do are typically:
– Reading large PDFs end-to-end
– Having conversations that span hours
– Using Claude Projects with too many knowledge files

What happens when you run out

When you hit the context limit, Claude can’t fit any more in. There are two things that might happen:

1. The platform truncates older messages. Claude.ai may silently drop the oldest parts of the conversation. You won’t be told. You’ll just notice Claude has forgotten something it knew earlier.

2. The platform refuses. Some interfaces will return an error if you try to send a message that pushes you over the limit.

Either way, the fix is the same: start a new conversation, and only carry forward what you actually need.

How to manage context

A few practical patterns:

For long PDFs. Don’t keep the whole PDF loaded across a long conversation. Get the summary you need, then start a new chat. Or use a Claude Project and put the PDF in the knowledge files so it’s available without being in the chat history.

For long workflows. Break them into phases. Start fresh for each phase. Carry forward only the outputs of the previous phase, not the whole conversation.

For Projects. Be ruthless about what goes in the knowledge files. Every file you add takes from the same 200,000 budget. Don’t include yesterday’s emails just because you can.

What this costs

Tokens are also how Anthropic prices the API. If you’re using Claude through claude.ai (the web interface), tokens are bundled into your Pro subscription — you pay ₹1,800/month and use Claude as much as you want, within fair-use limits.

If you’re using the API (typically via n8n or your own code), you pay per token. As of mid-2026, Claude Sonnet costs roughly $3 per million input tokens and $15 per million output tokens. That sounds abstract; in practice, a daily morning briefing using Claude API costs perhaps ₹2–5 per day. A heavy automation might run ₹500–2,000 per month.

The phrase “burning tokens” is real but usually overblown. For most business use, token cost is not the constraint. Engineering effort and decision-making are the constraints.

The thing nobody mentions

The 200,000-token limit is a ceiling, not a target.

Just because Claude can read 500 pages doesn’t mean you should make it. Claude’s quality of response degrades when context gets very long — it starts to miss things, especially in the middle of long inputs. A focused 20-page input often gets a sharper response than a sprawling 200-page input.

Use the context window when you need it. Don’t fill it just because it’s there.

Last reviewed: 7 May 2026