Claude will continue your novel. The chat window is what stops you

The people asking how to continue a novel with Claude are mostly already paying for it. They have a claude.ai tab, a Project with fourteen chapters uploaded, and a routine: scroll to find where the story stopped, ask for the next scene, copy the reply into a document somewhere else. It works, in the way that eating soup with a fork works. The model is not the bottleneck. The container is — and swapping the container is a twenty-minute job, most of which is waiting for a billing page to load.

Is Claude actually good at continuing fiction?

Good, with a documented failure pattern — and never the winner. In our 360-round blind benchmark (nine models, two novels, 20 consecutive rounds each, double-blind review), Claude Opus 4.8 placed fourth to fifth in both genres. That consistency across a fantasy epic and an ornate palace novel is itself unusual: most models swung wildly between the two books.

The reviewer notes explain both halves. On the palace novel Claude wrote the sharpest verbal sparring in the field — courtly needling, wit under etiquette — which is exactly what people mean when they say Claude has a feel for dialogue. But in both genres the long run turned volatile: heavy literary flourish early, a strong middle, then punctuation corruption and drifting names late. The reviewers called it the most volatile trajectory of the fantasy field. In practice that means Claude rewards a workflow where you can regenerate one wobbly segment or hand a chapter to a different model, and punishes one where forty chapters ride on a single unbroken run.

Worth noticing for a buyer: only two models held their band across both books — GLM 5.2 in the top three, Claude in the four-to-five slot — while every other entrant swung between podium and floor when the genre flipped. Steady-but-never-first is a legitimate thing to pay for. It just should not be bought blind, which is what the free-audition math later in this post is for.

Why isn’t claude.ai Projects enough for a long novel?

Projects will hold your book; they will not behave like one. A Project is a workspace with a knowledge base, and per Anthropic’s help center (as of July 2026), when that knowledge outgrows the context window, Claude switches to retrieval — pulling excerpts it judges relevant instead of reading everything. For question-answering that is a feature. For continuation it means you no longer know which parts of your book the model actually saw before it wrote the next scene.

The window itself is generous now: Anthropic’s published limits give Sonnet 5 chats a 1M-token window on paid plans, with Opus 4.8 and Sonnet 4.6 at 500K. A 300,000-word novel fits. What Projects still lack is everything around the window: there is no reading position, so you scroll chat history to find the story; a regenerated reply replaces the old one instead of sitting beside it; and the output accumulates as chat bubbles you have to copy somewhere else to read comfortably. None of that is a flaw in Claude. It is a chat product being used as a reader.

The honest carve-out: if you finished a book and occasionally want to ask “what if the ending went the other way”, Projects cost you nothing beyond the subscription you already pay. The API route below is for the other case — you are reading, and you want the story to keep going where you are standing.

How do you wire the Claude API into a phone reader?

Four steps, about twenty minutes end to end. One thing to know before you start: claude.ai subscriptions and API billing are separate tracks. A Pro or Max plan includes no API credit, so this route means creating billing on the developer console even if you already pay for the chat product.

Step one, create the key. Sign up at console.anthropic.com — it lands on the same developer console as platform.claude.com — and add a card or prepaid credits under Settings → Billing first, because a fresh key is rejected until billing exists. Then Settings → API keys → Create Key. The sk-ant- string is shown exactly once; copy it on the spot.

Step two, add it to a BYOK reader. In Foreverse that is Settings → AI models & services → Model providers → Anthropic, paste under API key, save. The preset host is https://api.anthropic.com/v1 on Anthropic’s native protocol, so nothing else needs configuring — and the same screen accepts any Anthropic-compatible relay endpoint if you use one. Run the built-in capability test: it makes one real call for a fraction of a cent and tells you immediately whether the key and host agree. The key is stored encrypted on the device and requests go straight to Anthropic; the BYOK writing page covers that architecture.

Step three, set the default text model under Settings → AI models & services → Default model config. Or skip it — the continuation form lets you pick per request, which is how you audition two Claude tiers against each other on your own chapters.

Step four, import the book and continue. Bring in the txt or epub, read to the paragraph where you want the story to grow, long-press it, tap the continue action, and give a one-line direction if you have one. The new text lands on a branch: the original is never modified, a bad take is a short branch you abandon, and the context for each request is assembled from the book automatically — recent chapters in, earlier AI continuations labeled as canon. That labeling matters more than it sounds; our prompting field notes measured a model going from 20/20 restart loops to zero on the strength of one canon sentence.

What does a Claude continuation cost?

Around three to seven cents at July 2026 list prices, depending on tier. Assume a typical request — 10,000 input tokens of context, 800 tokens out. Anthropic’s pricing page lists Claude Sonnet 5 at $2 per million input tokens and $10 per million output as introductory pricing through August 31, 2026 (then $3/$15): about $0.028 per continuation. Opus 4.8 at $5/$25 lands near $0.07. Haiku 4.5 at $1/$5 halves the Sonnet figure, though we have no blind fiction data on it. Prompt caching softens all of this — cache reads bill at a tenth of list input, and consecutive continuations reuse most of their context, so warm chains run cheaper than the cold math.

Caching deserves one concrete number because continuation is its best case. Every request re-sends mostly the same book context, and Anthropic bills cache hits at a tenth of base input — $0.20 per million tokens on Sonnet 5’s introductory table, $0.50 on Opus 4.8. If 9,000 of your 10,000 input tokens land on a warm cache, the input side of a Sonnet continuation drops from two cents to under half a cent. The app’s request log records cache hits per call, so this is a discount you can audit rather than take on faith.

For scale: a heavy 200,000-word ride — roughly 500 continuations — costs about $14 on Sonnet 5 introductory pricing, $35 on Opus 4.8 before cache savings, versus about $0.80 on DeepSeek V4 Flash, the budget model that won our fantasy table. Claude is the premium option and the premium is real, seventeen to forty times the cheapest credible alternative. It is also, in absolute terms, the price of two paperbacks for a whole alternate ending. The full cost ledger works the same arithmetic across every route.

What if you’d rather not open an API account?

Two Claude entries — Sonnet 4.6 and Opus 4.8 — sit in Foreverse’s metered official channel as of July 2026, which needs no Anthropic account at all: sign in, and the 5,000 starter credits (roughly 260 continuations at our measured average) work on Claude directly. The disclosure that keeps this paragraph honest: official-channel pricing is the model’s list price plus 50% as a service fee. Burn the grant first, and open the API account only if Claude earns a permanent seat in your rotation.

The twenty-minute version of all this: billing, key, paste, test, read. The part worth carrying away is smaller than a tutorial — Claude was never the thing failing at your novel. The fork was.