AI Novelist

What is AI Novelist?

AI Novelist (Japanese: AIのべりすと) is a browser-based Japanese prose continuation service from Bit192, Inc. You type an opening, press continue, and a Japanese-first model extends it in the voice already on the page. The manuscript is the prompt, not an instruction to an assistant.

Its operator calls it the largest public Japanese storywriting AI, trained from more than 2TB corpus — a scale claim, not a quality measure. Three properties separate it from a general writing assistant: the text the model reads is capped by membership tier; direction comes through in-text notation, not a settings panel; and the terms are unusually permissive on rights to generated text.

Core features

Membership-scaled context ceilings

How much manuscript the model sees is tied to your plan. AI can recognize up to 24,000 characters (with a free account), 48,000 characters (with a Voyager member), and even up to about 300,000 characters (with a Platinum member) long — a roughly twelvefold gap, which is why long-form consistency differs sharply across tiers. Context alone is not stability: to create a stable flow, you need to build a context of at least 40 lines, below which the model invents proper nouns.

In-text direction instead of an instruction panel

• Bracketed scene annotations. By enclosing the outline of the scene you want to write in [] (brackets) in the text, the model reads that line as stage direction, not prose, steering for about 30 lines to a maximum of 100 lines along the annotated content. (Compatible with SuperTrin and Damsel+RAY only).
• Comment markers. Lines beginning with @_ are comment lines and are not sent to the AI, keeping notes and cut branches in the file without spending context.
• Character Book with conditional commands. Using @commands, you can configure character book behavior in more detail — scoped by scan range, prioritised, gated on co-occurring names, or fired probabilistically.

A model roster with different jobs

Language depth is graded. For the Spiko-sama line the docs list Japanese as first language, English, Korean, Simplified Chinese, Traditional Chinese as second, and French, German, Spanish, Portugese as third. The legacy set is Japanese only and frozen: following models are discontinued and will not recieve future update.

Default, Solid and Colorful trade obedience against invention rather than stepping up in quality. Colorful routes to a sub-model chosen by membership level; those sub-models are specialized in generating natural novel continuations and are not designed to actively follow instructions.

Custom Style MODs

MOD is user-facing fine-tuning: you supply a corpus, the service trains an adapter, and the style sits between input and output. Official MODs cover light novella, romance, Sengoku historical and tabletop RPG replay. Input rules are hard: corpora must be plain text — pdf, Word and iWork formats are not supported — and the character code of the text used for the corpus only supports UTF-8 (Shift-JIS and EUC-JP are not supported).

Guide

From blank page to stable draft

1. Register with an email address and password. No name or other personal details are required, and registration is free.
2. Seed at least forty lines, writing the setting yourself or correcting a drafted one.
3. Correct rather than regenerate. The model imitates the page, mistakes included; line breaks, spaces, and symbols also influence AI's thinking.
4. Annotate the next beat with a bracketed line, which states what happens next without becoming text the model copies.
5. Force scene changes explicitly. Without a divider such as an asterisk row, the model keeps depicting the same scene at finer resolution.
6. Adjust repetition penalty when output loops. A high value trades instruction adherence for spontaneous development.

Character Book entries are conditional injections. @range 400 scans only the last 400 characters; @and Trin|Weirs fires only when those names are in range; @addbias shifts word likelihood, with bias between -30 and 30 recommended.

Use cases and worked examples

• Long-form continuation, with Spiko-sama Max named as the recommended model for long stories.
• Body text from a bullet plot. The Spiko-sama models were tuned to generate body text from bullet-point plots; feed one or two events at a time and enter the elements you want described in the next 300-1000 characters.
• Text games and chat-style play, via a Chat/Game GUI mode that adds an input field feeding the main text area.
• Poetry, lyrics and tanka. Derrida is the documented model, and a genre hint on the first line sharpens otherwise generic results.

Who is it for

Where it fits, and where it does not

• Japanese fiction writers who want their own voice extended, since the models imitate the register on the page rather than imposing a house style.
• Long-form authors on a paid tier, where a 300,000-character ceiling separates a tool that forgets your first act from one that does not.
• Creators needing clear commercial rights, and streamers, since streaming is permitted with no prior or subsequent permission.
• Poor fit: teams needing confidentiality, since token data sent to the AI cloud server is processed unencrypted.
• Poor fit: writers working mainly outside Japanese, since second- and third-language support is graded and the Japanese-only models carry none of it.
• Poor fit: under-15 users without a guardian, and anyone needing billing flexibility. Suspense and horror genres may generate content inappropriate for people under 15, who are asked to obtain guardian consent.

Alternatives and how to weigh them

Comparison material in the Japanese creator community frames this choice along two axes rather than as a ranking: Japanese-register fidelity versus structured setting management, and specialist tool versus general assistant. Those write-ups are editorial assessments, not benchmarks, so the useful output is the dimensions, not a verdict.

The dimensions that change a decision come from the product's own pages: whether your context ceiling covers your manuscript; whether you want direction through in-text notation or a panel; and whether the rights and confidentiality terms fit your publishing plan.

Platforms

The primary surface is the web application on the operator's domain, used in a desktop browser. No separate desktop or mobile client is documented; an independent Japanese directory lists it as a cloud service with a free tier and paid plans. Browser extensibility comes through third-party user scripts requiring Tampermonkey.

Developer API

The API carries an explicit scope limit: currently in limited alpha. We are not supporting large-scale public deployment (except personal client/app) yet. Treat it as fit for a personal tool, not public infrastructure.

Two request styles are documented. A text completion endpoint mirrors GUI continuation and exposes the full sampler surface — temperature, top-p, top-k, min-p, repetition penalty, banned words, logit bias, stop sequences. Models such as Spiko Ultra also work with an OpenAI-compatible v1 endpoint used in most CLIs and SDKs, with streaming support.

Limits are published per model. Most allow 300 requests per minute; legacy Damsel 20B allows 90. Maximum input runs from 2,400 tokens on Damsel 20B to 200,000 on Spiko Ultra. The documented rule of thumb for ordinary Japanese text is 1.5 characters per token, and context tokens plus tokens to generate must stay at or below the model's maximum sequence length.

Pricing

Tier names appear throughout the documentation as functional gates, not marketing labels. The membership pricing page requires sign-in, and this review could not read current monthly rates while signed out, so no monthly figures appear here.

Membership also changes which model answers in Colorful mode: for Spiko-sama-04 and Spiko Ultra the sub-model is Spiko-sama Max on Platinum and Spiko-sama-02 on Voyager. An independent Japanese directory summarises the paid effect as unlocking MOD use and a wider choice of language models.

The API is metered separately in a credit unit called Lumina rather than bundled into membership, priced per 1000 tokens and split between context and generation.

Two consequences follow. Spiko Ultra has the cheapest context rate of any current model despite being the largest, making generation the dominant cost. And MOD training consumes Lumina according to the number of training steps, so a custom style is a metered purchase.

Payment runs through two channels: a Stripe recurring subscription and a PayPal one-off payment. Stripe accepts major credit and debit cards, Apple Pay and Google Pay, and VISA prepaid cards; PayPal also accepts bank transfer.

Limitations

Rights, confidentiality and conduct

Generated text and images may be used freely for private or commercial purposes at the user's own risk, no attribution is required, and Bit192 states it holds no copyright whatsoever, including moral rights, over generated content. The same clause disclaims responsibility for consequences of sharing it: the permission is broad, the liability stays with you.

Confidentiality is the counterweight. Because token data sent to the cloud servers where the AI runs is processed without encryption, 100% confidentiality is not guaranteed, and the operator explicitly assumes no confidentiality obligation. Stored documents are encrypted with a key only the user knows, and the operator states it does not inspect them. Storage at rest and text in transit carry different promises, which matters most under a client NDA.

Four generation categories are prohibited outright: content judged to violate Japanese law, local law, public order or third-party rights; political content; content promoting affiliate schemes, multi-level marketing, cryptocurrency, NFTs or other investment products; and content judged to target racial or religious discrimination.

Billing, account and technical boundaries

Refund terms are restrictive and consistent across two pages. Users who delete an account mid-term are not refunded, and refunds for mid-term cancellation are not accepted; the help centre repeats that requests for personal reasons such as no longer using the service are declined. Cancellation does not cut access immediately: cancelling through the customer portal leaves premium features available until the paid term ends. One account rule matters before committing a long project: if no operation is performed on an account for 180 days or more, or if improper conduct occurs, Bit192 may delete or suspend it at its discretion, with no refund of fees already paid.

• MOD corpus size limits are documented inconsistently. The English help page states the MOD trainer reads a maximum of 1 million characters, truncating beyond that, while the Japanese manual states a maximum of 16MB per upload. The units and figures differ; this review reports the discrepancy rather than choosing one. On iPhone and iPad, uploads above 4MB may fail outright.
• Bracketed annotation does not cover the newest models. Documented compatibility is SuperTrin and Damsel+RAY only, so the technique most associated with the product does not transfer to Spiko-sama.
• Spiko-sama breaks compatibility with existing prompt assets: it is not compatible, and will not work correctly, with the officially supported Chat Prompts or most existing scripts.
• Advanced sampler controls are locked on the Spiko-sama line. Parameters are fixed and auto-adjusted except for a listed subset: randomness within 0.95 to 1.50, repetition penalty, typical P, output length and read-context length.
• Reasoning is not retained across turns, since only the most recent thinking block reaches the model, and the API is not cleared for production scale.

Privacy and data handling

Registration collects an email address and password only. Beyond that, the operator collects referrer data, IP address, server access logs and cookies or equivalent identifiers, used for service operation, traffic measurement and ad delivery. You can ask that collection or use of some or all of this information stop via the service's settings, but with a limit: for items whose collection is a precondition of running the service, it stops only once you withdraw from the service by the operator's prescribed method. MOD training carries a more protective statement: the uploaded corpus is deleted after training completes, and the MOD's unique ID is designed so the uploading user cannot be inferred.

FAQ

Q1. Can I sell a novel written with AI Novelist?

Yes. The terms permit free use of generated text and images for private or commercial purposes at your own risk, with no notation of service use required. Bit192 further states it holds no copyright at all, including moral rights, over generated content.

Q2. How much of my manuscript can the AI actually read?

It depends on membership, not the model alone: 24,000 characters on a free account, 48,000 for Voyager and about 300,000 for Platinum. Through the API the limits are given in tokens instead.

Q3. Which model should I use for a long novel?

The docs name Spiko-sama Max as the recommended model for long stories, on the grounds that it received continued post-training aimed at keeping the user's prose intact. Spiko Ultra is positioned for the longest context.

Q4. Is my writing kept private?

Partly. Documents saved to storage are encrypted with a key only you hold, but token data is processed unencrypted inside the cloud servers running the AI.

Q5. What happens to a corpus I upload for MOD training?

It is deleted from the server after training completes, and the MOD's unique ID is designed so the uploading user cannot be inferred from it.

Q6. Can I get a refund if I stop using it mid-month?

No. Fees already paid are not refunded when an account is deleted mid-term, and refunds for mid-term cancellation are not accepted. Duplicate charges are the stated exception.

Q7. Can I stream or make videos using it?

Yes, commercially and non-commercially, with no prior or subsequent permission required, and monetisation through ads or merchandise links is within the permitted range.

Q8. Is the API ready for a public product?

Not by its own scope statement: it is published as in limited alpha with large-scale public deployment unsupported. Rate limits are 300 requests per minute on most models.