Continuing a novel with DeepSeek means getting past three walls
The chat box at chat.deepseek.com can't hold a full novel; the API lives on a separate platform (register, create a key, top up); and a raw key still needs a reader that assembles context from the whole book. This teardown takes the walls in order, with July 2026 list prices ($0.14 per million input tokens on V4 Flash, cache hits at 1/50th), the July 24 model-name retirement, and where DeepSeek actually placed in our blind continuation benchmarks.
Anyone who has tried remembers the moment: select all on a two-million-character webnovel txt, copy, paste into the box at chat.deepseek.com. The browser stutters, and what actually lands is a short tail of the book — everything before it silently gone. Most searches for “continue a novel with DeepSeek” start at that moment.
First, the reassurance: this road goes through. DeepSeek models took first place in both genres of our blind continuation benchmarks (the numbers are below). But between the chat box and actually continuing a book stand three walls.
Wall one: why the chat box can’t hold a book
A chat box is a container built for conversation. A single paste holds only so much; a new session forgets every character and plot thread you established; and even if you hand-feed it a few chapters, the next continuation makes you curate all over again. What’s blocking you isn’t model capability. It’s container shape.
The spec-sheet rebuttal writes itself: DeepSeek advertises a 1M-token context window with 384K max output — surely a whole book fits? It fits. The accounting doesn’t. Continuation is a repeated act, and stuffing the full text into context means every single segment re-bills the entire book as input, nearly all of it irrelevant to the scene at hand. What works is sending only what matters each time: character sheets, planted foreshadowing, the recent stretch of plot. Doing that curation by hand in a chat window is the job nobody does a fourth time.
There are structural potholes beyond cost — generation that overwrites instead of branching, no way back when a direction sours — that the branching post takes apart, so this wall ends on one sentence: keep the model and change the container around it.
Wall two: opening the API
Three steps, all in a browser. Register at platform.deepseek.com — note this is the developer platform, a different door from the chatbot. Create a key on the console’s API Keys page. Top up on the billing page. Ten minutes end to end on a good day.
Details worth knowing before they bite. The generated key (an sk- string) displays in full exactly once; after the dialog closes the platform keeps only a hash, so copy it immediately or delete and re-create later. The API is prepaid, and at zero balance calls fail with a 402. And before charging anything, open the usage page: the official deduction rules include a granted-balance mechanism spent ahead of your own funds, though how much a new account gets varies with promotions — first-half-2026 tutorials circulated anything from one to five million tokens, which is exactly why you should read your own number instead of budgeting on a forum post.
Prices, per the official pricing page (checked 2026-07-22): deepseek-v4-flash at $0.14 per million input tokens ($0.0028 on cache hit) and $0.28 per million output; deepseek-v4-pro at $0.435 in and $0.87 out. In practice, a continuation with 10,000 tokens in and 800 out costs about a sixth of a cent on Flash.
That parenthetical hit price deserves a second look. Continuation is precisely the cache-friendly workload — every request re-carries the same book’s preceding text, and the repeated part bills at one-fiftieth of the miss price once cached. The caching post measures what that does to a monthly bill.
Two dated notes from this July. DeepSeek has announced peak-hour pricing alongside the V4 official release — doubled rates inside two Beijing-time day windows — though as of our July 22 check the international pricing page still lists one flat rate; the sunset calendar tracks the switch. And the legacy names deepseek-chat and deepseek-reasoner retire July 24 at 15:59 UTC. New configs should say deepseek-v4-flash or deepseek-v4-pro; old tutorials copied verbatim start erroring out the moment that timestamp passes.
Wall three: where the key goes so continuation actually works
Into a BYOK reader — an app where you bring your own key, and which does the two jobs the chat box couldn’t: guard the key, and assemble each continuation’s context from the whole book automatically. Requests travel from the phone straight to api.deepseek.com, no third-party server in the path, at the list prices above with zero markup.
Concretely, in Foreverse (Android, on Google Play): Settings → AI models & services → Model providers, find DeepSeek among the 62 preconfigured providers, paste the sk- key, save. The provider page has a capability test — one real call, cost measured in fractions of a cent — that settles on the spot whether the key works. Then either set V4 Flash as the default text model or pick a model per request on the continuation form.
The rest happens inside the book. Import the txt or epub, read to where the story needs a push, long-press to select a paragraph, tap continue, type a one-line direction. The AI picks up from that exact spot; the new text lands on a branch beside the original, and a direction that sours gets deleted without the source text ever changing.
One alternative, to save you an account you might not need: skip the DeepSeek registration entirely and use Foreverse’s official metered channel — sign in, and the 5,000 signup credits cover roughly 260 continuations, with both DeepSeek tiers in the catalog. The trade is a 50% service margin on list price. Heavy users come out ahead opening their own key; curious ones can spend the grant first and decide after.
Is DeepSeek worth pointing at fiction at all?
Genre decides, but both genre crowns in our benchmarks belong to DeepSeek. In the six-model fantasy run — each model continuing an 8.9-million-character epic for 20 consecutive rounds, outputs double-blind reviewed — V4 Flash won outright, with the reviewers calling it the only system that read like a continuation chapter of the original. In the nine-model run on an ornate court novel, the top tier went to its sibling V4 Pro.
The fun part is that the two tables contradict each other: fantasy champion Flash fell to 6th–7th on the court novel, its plain-spoken, fast-moving instincts useless against formal period diction. Match the tier to the book — Flash for fast-paced fantasy, Pro for formal period prose — before asking anything else. Full rankings, review quotes, and the protocol’s limits live in the model-picking page and the 120-round fantasy arena.
Three walls down, the ledger is small: ten minutes to open the platform, two to paste a key, a sixth of a cent per segment. The two million characters that wouldn’t fit in the chat box now sit whole in a reader — and which slices of them ride along with each request stopped being your problem.
FAQ
Does DeepSeek's granted balance cover continuation writing?
It can, but don't budget on hearsay. The official deduction rules confirm a granted-balance mechanism that is spent before your topped-up funds; how much a new account receives, and for how long, shifts with promotions — tutorials in the first half of 2026 circulated figures anywhere from one to five million tokens. The reliable move is to open the platform's usage page after registering and read the actual number. When granted balance runs dry and nothing is topped up, API calls return a 402.
Is DeepSeek cheap enough for everyday fiction continuation?
At July 2026 international list prices, yes, almost absurdly so. V4 Flash charges $0.14 per million input tokens on a cache miss, $0.0028 on a hit, and $0.28 per million output. A typical continuation — around 10,000 input tokens and 800 out — lands near $0.0016. Continuation is also cache-friendly by nature, since each request re-sends mostly the same book context; the hit price is one-fiftieth of the miss price, which our caching write-up unpacks in full.
Old tutorials say to use deepseek-chat — why does that fail now?
The official pricing page carries the notice: deepseek-chat and deepseek-reasoner retire on July 24, 2026 at 15:59 UTC. They were compatibility aliases for deepseek-v4-flash's non-thinking and thinking modes, so new configurations should name deepseek-v4-flash or deepseek-v4-pro directly. Any config copied from an older tutorial keeps working until that timestamp and then errors out — a one-line model-name fix, but a confusing failure if you don't know the date.
Is DeepSeek actually good at continuing fiction, or just cheap?
Both, with genre fine print. In our six-model fantasy benchmark (20 consecutive rounds per model, double-blind review), V4 Flash won outright — reviewers called it the only system that read like a continuation chapter of the original. In the nine-model ornate-court benchmark, its sibling V4 Pro took the top tier while Flash fell to 6th–7th. Same vendor, two tiers, two genres: pick the tier that matches the book, not the brand.
Questions or ideas? Join our Discord →