Every flagship ships 1M context now. Your novel fits. That's not the question

The spec sheets converged this summer. Claude Fable 5 launched June 9 with a million-token window as the default, no beta header, no long-context surcharge. DeepSeek’s V4 line carries 1M across all official services. GPT-5.6 reads 1.05M. And on July 16, 2026 — two days before this post — Moonshot shipped Kimi K3 with a 1,048,576-token window at one flat rate. Google, for the record, has been past this milestone since 2024, when Gemini 1.5 Pro opened a 2M window to all developers, and this spring’s coverage reports 2M again on the newest Ultra tier. The marketing pitch for readers writes itself: your entire serial finally fits in one request, so the AI can finally write it properly.

We happen to have run the control experiment for that pitch before the race got loud. The short answer: fitting is not learning. The rest of this post is the data, and the bill.

The arms race, dated and sourced

By the arithmetic in our context-window explainer — which answers the beginner question this page skips, namely what a window even is — a 500k-word serial runs roughly 625k to 830k tokens. Under every row of that table, it fits. Three years ago it didn’t fit one-fifth of the way in. That flip is real, and it’s worth taking seriously before we take it apart.

The experiment the pitch never runs

If window size were the binding constraint on faithful continuation, more original prose in context should mean output more like the original. So we measured exactly that, before our multi-round benchmarks: a single-shot control across five context tiers, from 4k tokens of original prose up to 200k — at the top tier, an entire 270k-character fantasy novel, crammed in whole. The full write-up is currently published in Chinese; the numbers below are from that run.

The no-instruction arm is the one that kills the pitch. Handed the whole book and told nothing about style, the model wrote boilerplate similes at ten times the original author’s density — the Chinese equivalents of “as if” and “a wisp of” on every page, in a novel whose author uses about one simile per thousand characters. All that prose sat in the window, and the model went on writing in its own default voice. It saw everything. It imitated itself.

The instruction arm is the counter-shot. One explicit sentence — treat the original passages as the style baseline — and the largest window didn’t drown it out: the direction held at 200k just as it did at 4k. Across models, tier-to-tier differences in the single-shot setting were mild either way. What separated faithful from unfaithful was never how much of the book the model saw. It was whether anyone told it what the material was for. Volume is a substrate; the baseline declaration is the mechanism. The second one is free.

The bill for cramming

Now the honest ledger, using the friendliest pricing in the table. Kimi K3 charges $3.00 per million input tokens, flat across the window — no long-context tier, which is genuinely the least hostile billing design of the bunch. A cold request filling all 1,048,576 tokens: about $3.15 of input, before the model writes a word. Your 700k-token serial: about $2.10 per press of the continue button. Twenty presses over an evening of reading is roughly $42 of input, re-buying the same unchanged chapters twenty times.

Caching is the standard rebuttal, and it deserves a fair hearing. K3 prices cached input at $0.30 per million, so a warm re-send of that serial drops to about $0.21 — a 90% discount, automatic on stable prefixes. The catch is the word stable: the discount holds only while the prompt prefix stays byte-identical, and continuation is precisely the workload where you edit a paragraph, regenerate a scene, branch at a choice point. Every upstream change resets the meter to $3.00 territory for everything after it. Caching rewards append-only reading; it does not absolve cramming. GPT-5.6 Terra runs the same errand at $2.50 per million — the explainer post did that arithmetic and landed on $1.75 per press.

What the big windows are actually good for

Credit where due. The 1M generation removed a real failure: on 200k-class windows, a long serial literally could not be sent, and the oldest chapters were amputated unread. That failure is gone, and workflows that need one huge read — summarize this whole book, find every scene with this character — got materially better. What the big windows did not fix is retrieval within the window: OpenAI’s own long-context eval on the GPT-5.6 launch page reports 73.8% on retrieving 8 planted items at 512K-1M depth. That is the vendor grading its own product, at the depth where chapter 3 of your serial lives. And nothing at any window size fixes cross-request persistence, which remains zero — next press, you send it all again.

Feed the scene, not the shelf

Our position hasn’t changed since before the arms race: the window is a reading budget, and the winning move is spending it on relevance instead of volume. Always-on facts — who the protagonist is, how the magic works — travel in full with every request. Everything situational becomes lorebook entries with trigger keywords, entering the request only when the current passage mentions the Northern March, costing zero tokens the rest of the time. The 360 rounds of continuation behind our style benchmark and our failure-mode study all ran inside fixed 16k-token windows — scene, triggered entries, recent prose — and across 360 rounds, none of the failures we documented would have been fixed by the millionth token. The restart loops and mid-run freezes happened inside contexts a 2023 model could hold.

The world-organizer page shows what that structure looks like in Foreverse. The honest close is an admission: we’d rather sell you a 16k-token habit than a million-token bill, and on our own benchmark data the habit also writes better fiction.