One Prompt AI Book: can Gemini 2.5 write a book in a single prompt?

Among all the hotly debated parameters in the latest AI models, maximum output tokens seems curiously under-discussed. Most models have had surprisingly limited output ceilings. So when Gemini 2.5 launched with a claimed 65,536 max output tokens, a fourfold increase from Gemini 2.0, I had one immediate question: is that enough to write an entire book?

Putting it in book terms

65,000 tokens translates to roughly 48,750 words, which is about 195 double-spaced A4 pages or 139 paperback pages. For context, that's approximately the length of The Adventures of Sherlock Holmes. So theoretically, yes, we're in book territory.

The test

I wrote a prompt requesting 38-40,000 words of fiction. The plot was loosely based on my own emigration story, deliberately made a bit ridiculous. I tested it through Google AI Studio, a Python script, and also through Anthropic's Sonnet 3.7 for comparison.

The actual results

This is where it gets interesting. Both Gemini methods produced only about a quarter of the claimed maximum: roughly 11,500 words via AI Studio and 10,500 via the script. That's about 23% of the theoretical ceiling. Anthropic's Sonnet 3.7, however, produced 27,655 words, over 56% of the claimed maximum and more than double what Gemini managed.

The token-to-word ratios were fairly consistent across methods (1.26 to 1.36 tokens per word), so the difference really comes down to how much continuous output each model is willing to produce before deciding it's "done."

Key takeaways

Claimed max output tokens and actual output are very different things. Gemini 2.5 delivered roughly a quarter of its advertised capacity, while Anthropic delivered just over half. Neither hit their ceiling. But the trajectory is clear: we're moving toward models that can sustain coherent, long-form generation, and that opens up genuinely new creative possibilities.

All the prompts, scripts, output text, and detailed analysis charts are available on GitHub.