#tHow Many Tokens?

← Back to counter

How many tokens are in a book?

The short answer

A standard novel is 80,000 to 130,000 tokens. Most novels fit comfortably in modern long-context models. GPT-5's 400k window or Claude Sonnet 4.5's 200k window holds nearly any single book.

Rough heuristic: ~1.3 tokens per English word for OpenAI's tokenizer, slightly higher for Claude's (which uses ~35% more tokens for the same text after the Opus 4.8 update). So a 70,000-word novel ≈ 91,000 tokens on GPT, ~120,000 on Claude.

Worked examples from real books

These numbers were generated using OpenAI's tiktoken (cl100k_base) for GPT-4-class models. Claude counts run 30-35% higher.

BookApprox wordsGPT tokensClaude tokens
*The Great Gatsby*47,000~62,000~84,000
*Of Mice and Men*30,000~40,000~54,000
*To Kill a Mockingbird*100,000~135,000~182,000
*Pride and Prejudice*122,000~165,000~223,000
*1984*89,000~120,000~162,000
*Brave New World*64,000~85,000~115,000
*Lord of the Rings* (one volume, *Fellowship*)187,000~245,000~330,000
*Harry Potter and the Sorcerer's Stone*77,000~102,000~138,000
*The Bible* (King James)783,000~1,050,000~1,420,000
*War and Peace*561,000~750,000~1,015,000
Average academic textbook200,000~270,000~365,000

Will it fit?

This is the question most people are actually asking. Here's where common books fit in 2026's context windows:

ModelContext windowFits
Claude Haiku 3.5200,000Most novels; up to Pride & Prejudice on GPT counts, but Claude counts push longer novels over
Claude Sonnet 4.5 / Opus 4.8200,000Same as Haiku
GPT-4o / GPT-4o-mini128,000Most novels under ~95k words
GPT-5 / GPT-5 family400,000Single-volume LOTR, *War and Peace* in ~2 chunks
GPT-5 Pro (1M variant)1,000,000Entire Bible in one call; *War and Peace* with room to spare
Gemini 2.5 Pro2,000,000Multi-volume works in a single context
Gemini 3.1 Pro2,000,000Same

If you're working with a single average-length novel, anything from Claude Haiku upward will fit it. If you're working with reference books, multi-volume series, or full academic textbooks, Gemini Pro's 2M-token window is the only single-shot option in 2026.

Cost to send a whole book

Sending an entire book to a model costs money in proportion to its tokens. Real examples:

A 100k-token novel as input, with a 2,000-token output:

ModelInput costOutput costTotal per call
Claude Haiku 3.5 ($0.80/$4)$0.080$0.008$0.088
GPT-4o-mini ($0.15/$0.60)$0.015$0.001$0.016
Claude Sonnet 4.5 ($3/$15)$0.300$0.030$0.330
GPT-4o ($2.50/$10)$0.250$0.020$0.270
GPT-5 ($1.25/$10)$0.125$0.020$0.145
Claude Opus 4.8 ($5/$25)$0.500$0.050$0.550
GPT-5 Pro ($30/$120)$3.000$0.240$3.240

For analyzing books at scale, GPT-4o-mini at $0.016 per book is hard to beat. For careful single-book reasoning, GPT-5 at $0.145 per call is the sweet spot. For the highest-quality literary analysis, Opus 4.8 is the premium tier.

When you should chunk instead

Even when a book fits in a model's context window, sending the whole thing in one call isn't always optimal:

Cost. Most analysis questions don't need every word. A focused 20k-token chunk often gets better answers cheaper than the full 100k context.

Attention drift. Models perform worse at recall in the middle of very long contexts ("lost in the middle" effect). For specific lookups, retrieving relevant passages with embeddings and sending those is often more accurate than dumping the whole book.

Output quality. Long-context tasks tend to produce summaries-of-summaries; chunked tasks with iterative reasoning produce more granular work.

Rule of thumb: use full-book context for synthesis ("compare the themes of chapters 4 and 17"), use chunked retrieval for lookup ("what does the character X say about Y?").

How to count your own book

Drop the text into our GPT-5 counter or Claude counter for the exact token count. For very large books, split into ~50k-word chunks; both tokenizers are linear so the chunked counts sum correctly to the whole-book count.

If you're working programmatically, run tiktoken for OpenAI counts and the official count_tokens endpoint for Anthropic counts. The provider-specific counters on this site call those endpoints under the hood, so the result you see is what you'll pay for.

Open the calculator →

Try this on every model

Try the live counter →