NovelAI is a browser-based subscription service from Anlatan that pairs the NovelAI Diffusion V5 anime image model with in-house storytelling language models, giving writers and illustrators private generation, tag-level prompt control and unlimited text output from $10 a month.
Used this tool? Rate it
Used this tool? Rate it
NovelAI is a browser-delivered subscription service operated by Anlatan Inc., a Delaware-registered company, that combines an anime-focused image generation stack with a family of in-house storytelling language models. The homepage frames the offer plainly — "Create High Quality AI Anime Art & Stories" — and the two halves genuinely share one account, one subscription and one billing cycle rather than being sold as separate products. Writers who want a long-context model that keeps track of a novel's cast, and illustrators who want tag-level control over anime characters, subscribe to the same three tiers.
The service is unusual among AI creative tools in two respects that shape almost everything else about it. First, the image models from the V4 generation onward are described in the official documentation as "completely original image generation models, trained from the ground up without relying on a public base model like Stable Diffusion" — a departure from the platform's own 2022 origins, when its first image model was a heavily modified Stable Diffusion derivative. Second, generation is private by design: the FAQ states that image generations are session-based, that nothing is stored server-side, and that requests, uploads, edits and generations are not logged. That privacy posture is a genuine differentiator, and it also explains why the platform ships almost no content-side guardrails on private work.
The current flagship image model, NovelAI Diffusion V5, was announced on 21 August 2026 in both Curated and Full variants, adding true alpha transparency, multi-language text rendering, single-generation comic pages and character positioning for up to 22 distinct characters. On the text side the current flagship is Xialong, a 355-billion-parameter storytelling finetune of ZhipuAI's GLM-4.6 with a 28,672-token context window. Pricing starts at $10 a month; there is a one-time free trial but no permanently free tier.
The image documentation lists six selectable models, and the choice is not cosmetic — each behaves differently and is meant to be picked according to what you are trying to make. The two axes that matter are generation (V5, V4.5, V4) and dataset curation (Full versus Curated). Full models are trained on what the documentation calls "a more comprehensive, albeit less curated, dataset," while Curated models use "a somewhat smaller training set of cleaner and more focused images." Full is the choice when you need something specific that Curated cannot reach; Curated is the safer general-purpose default.
V5, released 21 August 2026, is the largest of these by a wide margin. Anlatan's own subscription announcement puts its architecture at 2.5 times the size of V4.5, which is the stated reason for the compute constraints discussed later in this article. The headline capabilities are sharper natural language understanding, expanded multi-language prompting, reworked character positioning supporting up to 22 distinct characters, higher image fidelity, true alpha transparency output, and a new "Max" enhancement level for upscaling.
One practical note that catches new users: the V4 and V4.5 models use the T5 tokenizer, and the documentation warns that "most Unicode characters (e.g. colorful emoji or Japanese characters) are not supported by the model as part of prompts." If you write prompts in Japanese out of habit because the aesthetic is anime, the model will not read them the way you expect on those generations.
NovelAI's prompt system is closer to a Danbooru-style tag vocabulary than to the conversational prompting of general-purpose image models, though V4 onward added genuine natural language understanding on top. In practice you can mix registers — a natural-language scene description plus a run of comma-separated tags for hair, eyes, framing and pose. The prompt token budget is roughly 512 T5 tokens covering the base prompt and all character prompts combined, which is generous but not unlimited once you are describing several characters in detail.
Emphasis syntax lets you strengthen or weaken individual concepts rather than rewriting the whole prompt, and negative numerical emphasis offers targeted removal of a concept that a blanket negative prompt would over-suppress. Sampler, steps, guidance and seed are all exposed directly in the interface, so a result you like can be reproduced or varied deliberately rather than by chance.
This is the feature most responsible for NovelAI's reputation among illustration users. Rather than describing two people in one prompt and hoping the model does not blend their attributes, you write a separate prompt block per character. The homepage demonstrates the pattern with two distinct character prompts — one for a girl with green hair and star-shaped pupils in a red dress, another for a woman with orange hair and a black jacket holding an ice cream cone — positioned side by side. V4 introduced this as a first-class capability and V5 extends the positioning system to as many as 22 characters in one image.
Attribute bleeding between characters is the single hardest failure mode in multi-subject anime generation, and a per-character prompt structure addresses it at the architecture level rather than through prompt-engineering tricks.
The Undesired Content field is where you list what the model should avoid, and it sits as a tab alongside the main prompt. NovelAI ships presets that are injected automatically, and it is worth being precise about what those presets contain, because the name invites a misreading. The Heavy V4.5 Full preset, for example, injects: lowres, artistic error, film grain, scan artifacts, worst quality, bad quality, jpeg artifacts, very displeasing, chromatic aberration, dithering, halftone, screentone, multiple views, logo, too many watermarks, negative space, blank page.
Every one of those terms is a quality descriptor. The documentation frames the purpose the same way: injecting the presets means "the AI avoids some general undesired aspects and produces higher-quality content." There is a Human Focus variant that adds anatomy terms such as bad anatomy and mismatched pupils, and a Furry Focus variant for the furry and kemono datasets. There is also a "No Default UC" option that turns the presets off entirely. This is an image-quality system, not a content-safety system, and users should not mistake it for one.
Beyond text-to-image, the platform carries a full editing chain. Image2Image changes details while following an existing composition, and Enhance raises the level of detail in an existing generation. Vibe Transfer extracts the aesthetic qualities of a reference image and applies them as direct visual inspiration to new generations — useful for holding a consistent look across a series without maintaining an elaborate style prompt. Inpaint lets you brush over a region to fix an error or prompt for something different in that area alone.
The Director Tools group handles post-processing: Remove BG, Line Art, Sketch, Colorize, Emotion and Declutter. These operate on a finished generation and are the kind of small utilities that normally require a separate editor, so having them adjacent to the generation history saves a meaningful amount of round-tripping.
The text side has been running since June 2021 and has its own model lineage, each named after a mythological figure. Xialong, released on 31 March, is the current flagship: a 355-billion-parameter finetune of ZhipuAI's GLM-4.6 on NovelAI's storytelling dataset, and the largest finetune the team has produced. It offers 28,672 tokens of context on the website plus an 8,192-token rollover window, meaning usable context in the story editor actually floats between 28,672 and 36,864 tokens as a story grows before being trimmed back down. The documentation notes a deliberate design property: unlike models trained with reinforcement learning, Xialong "provides genuinely different outcomes on retries and improved name variety" — which matters a great deal if your workflow involves rerolling a paragraph until it lands.
The earlier models remain instructive about the platform's trajectory. Kayra is a 13-billion-parameter in-house model trained on roughly 1.6 trillion tokens with an 8,192-token context, released 28 July 2023. Clio is a 3-billion-parameter sibling from May 2023 with the same context window. Erato, from September 2024, is built on Llama 3 70B Base with continued training on NovelAI's Nerdstash pretraining data. Opus subscribers additionally retain access to Xialong plus the older experimental models Krake and Llama 3 Erato.
Two toggles deserve a mention because they are easy to miss. Furry mode prepends a fur dataset, tag to your prompt, restoring the behaviour of the platform's earlier dedicated furry models within the V4.5 and V5 architecture. Separately, placing background dataset, at the start of a prompt shifts output toward landscapes, animal portraits and still lifes rendered in a photographic style, without people in them. That second option is a quiet answer to the common complaint that anime-tuned models cannot produce a plain environment plate.
Text-to-speech is bundled rather than sold separately: the free trial includes 100 generations and every paid tier from Tablet upward is unlimited. NovelAI Explore is the platform's public gallery, and it is entirely opt-in — nothing is shared unless you upload it deliberately. Every post goes through manual review before it appears, and the upload guidelines prohibit NSFW material, copyright, character or artist tags, and advertising. Images produced with Character Reference, Image-to-Image or Inpainting cannot be posted at all; the FAQ explains these are "harder to moderate and easier to misuse."
The most common workflow is designing a character and then holding that design steady across many images. Between seed control, Vibe Transfer for aesthetic consistency and per-character prompt blocks, this is achievable without training a custom model — which matters because the platform does not currently plan to let users train their own image modules. Tabletop groups, webcomic authors and original-character artists use this to build a visual bible before committing to finished art. For this audience the appeal of an AI anime art generator is precisely that the style is native rather than coaxed out of a general model.
Because text and image generation live in the same subscription, a writer drafting a light novel can generate scene illustrations without a second vendor relationship. The 28,672-token context on Xialong is large enough to hold a substantial amount of established setting, and the Lorebook mechanism injects relevant entries only when triggered, which keeps context spend proportional to what a scene actually needs.
V5's headline addition of full comic pages in a single generation changes what is feasible here. Previously, laying out a page meant generating panels separately and assembling them by hand; combined with the multi-language text rendering improvements, single-generation page output is a meaningful workflow shift for anyone producing sequential art. Expect to still refine panels with Inpaint — a full page generated in one pass will rarely be correct in every panel.
Text Adventure mode has its own interface and command mechanics. Xialong handles it with a specific caveat documented by the team: in Text Adventure mode user inputs are sent as instructions, and if you run adventures outside the dedicated mode you should use the { instruct syntax rather than the traditional > prefix. Anyone migrating an older adventure workflow to Xialong needs to know this or outputs will drift.
The Director Tools — particularly Remove BG and Line Art — make NovelAI usable as an upstream source for pipelines that finish elsewhere. Transparent-background output from V5 is directly importable into a layout or animation tool, and line art extraction gives a base for manual colouring. Under the platform's stated terms you retain ownership of what you generate, which makes this kind of downstream use legally straightforward as far as the vendor is concerned, though copyright questions about the output itself remain your responsibility.
Because nothing is logged or stored server-side, NovelAI is used heavily for work people do not want on a vendor's servers — unpublished manuscripts, commissioned character concepts under NDA, or simply material a user considers private. The trade-off is real and worth stating plainly: nothing is stored means nothing is recoverable, and a refresh of the image page empties the history queue.
Registration is at the main site and requires email verification before the free trial allowances unlock. The Paper free trial gives you 30 image generations up to 1024x1024 resolution, 50 text generations and 100 text-to-speech generations. Everything you configure and every story you write during the trial carries over if you subscribe, so there is no cost to setting things up properly from the start.
This is the step most new users skip. Selecting V5 Curated versus V5 Full changes what the model can reach, and the description of the currently selected model is displayed directly above the prompt box — click it to switch. If you are generating on stream or in a shared context, the documentation explicitly recommends Curated. If a specific concept is not landing, Full is the model with the broader dataset.
Start with subject and framing, then add appearance tags, then environment, then style. Adding everything at once makes it impossible to tell which term caused a problem. Use the dice button — the Random Prompt generator — when you want a starting point rather than a blank field.
The default settings are a reasonable starting point, and the platform's own guidance for saving usage is instructive: because cost scales with the pixel count of the chosen resolution, lowering resolution saves more than lowering step count. Keep steps at or below 28 and generate one image at a time if you are on Opus and want to stay inside the free-generation conditions.
If the composition is right but a detail is wrong, use Inpaint rather than regenerating. If the composition is wrong but the aesthetic is right, use Image2Image with adjusted strength. If a concept keeps appearing that you do not want, add it to Undesired Content or use negative numerical emphasis for a more surgical removal. The documentation offers a genuinely useful lateral example: the tag freckles may create artifacts, and adding tattoo to Undesired Content can remove the artifacts while keeping the freckles.
This cannot be overstated. Generations are session-based and are lost on refresh or when you leave the image generation page. There is a button at the bottom of the history sidebar that downloads the whole session as a .zip file — use it before you close the tab, not after.
For text projects, invest early in Memory, Author's Note and the Lorebook rather than relying on raw context. Lorebook entries activate on keywords, which means a large body of established setting costs context only when relevant. Stories export as .story and .scenario JSON files including all settings, Memory, Author's Note, Lorebook entries and modules, and there is a "Download All Stories" button under User Settings > Account for bulk backup.
Start with the recommended presets enabled, because they demonstrably improve output quality, but understand what they are doing. If your style intentionally involves screentone or halftone — plausible for manga work — the Heavy preset is actively fighting you, since both terms appear in it. Switch to Light, or to No Default UC and build your own list.
If a prompt produces messy results on Full, running the same prompt on Curated often reveals whether the problem is your prompt or the dataset. Curated's cleaner training set tends to produce more predictable compositions, so a prompt that works on Curated and fails on Full is usually a weighting problem rather than a description problem.
If you are on Opus, the four conditions for Anlas-free generation are worth internalising: one image at a time, no base image, at least Normal size, and 28 steps or fewer. Batch generation, image-to-image and high step counts all leave the free lane. Many users burn Anlas without realising it simply by leaving batch count above one.
Because nothing is stored server-side, a reproducible result depends entirely on you recording the seed alongside the prompt and settings. Downloaded images carry generation metadata, but if you export in a format that strips it, the recipe is gone. Check your download format setting before a long session.
Given the T5 tokenizer's handling of non-Latin characters on V4 and V4.5, English tags are the reliable path. This is counterintuitive for a product built around anime, and it is the kind of detail that produces silently mediocre results rather than obvious errors.
The terms give you a 45-day window to export content after a subscription ends, after which Anlatan reserves the right to delete it. Content is encrypted with a key derived from your password and email, and the terms are explicit that a lost password means unrecoverable content. Regular local exports are not optional for serious projects.
If you find yourself writing a 40-token style preamble on every prompt, that is a signal to use Vibe Transfer with a reference image instead. It holds aesthetic consistency more reliably than repeated style tags, and it frees prompt budget for the subject.
Anime and manga illustrators are the clearest fit. The tag vocabulary, the multi-character prompting, the six-model selection and the inpainting chain are all built around this style specifically rather than as one option among many. Anyone whose output is predominantly anime-styled will get more per dollar here than from a generalist model where anime is a prompt modifier.
Long-form fiction writers, particularly those working in serialised or web-novel formats, are the other core audience. The combination of a 28,672-token context, a Lorebook that scales with a story's complexity, and unlimited text generation from the $10 tier makes the economics predictable in a way that per-token API pricing does not.
Tabletop and interactive fiction creators get value from Text Adventure mode plus character art in the same subscription — a combination that would otherwise mean two vendors.
Privacy-sensitive users — writers under NDA, people developing unpublished IP, or anyone uncomfortable with prompt logging — are served by the no-logging, no-storage architecture in a way that few competitors match.
Who should look elsewhere: anyone needing photorealistic output, since the models are anime-tuned by design; teams needing collaboration features, shared workspaces or admin controls, none of which are part of the product; developers wanting a documented public API for third-party integration; and anyone requiring a native mobile app rather than a progressive web app. Users who need a permanently free tier will also be disappointed, since the free allowance is a one-time trial.
NovelAI is delivered entirely through the browser. The FAQ states that any computer or mobile device with an up-to-date browser can run it, given a stable internet connection, and that the website is additionally available as a Progressive Web Application installable from the site itself through a mobile browser. That PWA route is the official mobile experience.
There is no official native app. A search of the Apple App Store returns several products with confusingly similar names — "Novel AI - AI Story Generator," "NovelAI | AI Story Generator," "Novel AI : AI Story Generator" — published by unrelated individual developers rather than by Anlatan. These are separate products that happen to share the name, and their ratings, content ratings and reviews say nothing about the service at novelai.net. Users searching app stores should be aware they will not find the real thing there.
Infrastructure runs on Anlatan's own API, with the trained and finetuned models hosted on machines provided by CoreWeave. Payments are processed through Chargebee, and support runs by email through support@novelai.net with a separate corporate contact at contact@anlatan.ai. A public status page is maintained at status.novelai.net.
There are three monthly subscription tiers plus a one-time free trial. Subscriptions renew exactly 30 days after payment.
Paper (free trial) requires an email-verified account and provides 30 image generations up to 1024x1024, 50 text generations and 100 text-to-speech generations. Settings and stories created during the trial carry over on subscribing. This is a trial, not a free tier — the allowances do not refresh monthly.
Tablet — $10 USD per month. Unlimited text and text-to-speech generation for the duration of the subscription, plus 1,000 subscription Anlas per month for image generation.
Scroll — $15 USD per month. Everything in Tablet with a larger context size for storytelling, plus the same 1,000 subscription Anlas.
Opus — $25 USD per month. The largest context at 28,672 tokens, exclusive access to Xialong along with the older experimental Krake and Llama 3 Erato models, 10,000 subscription Anlas monthly, and Anlas-free image generation subject to four conditions: one image at a time, no base image, at least Normal size, and a step value of 28 or lower.
All three tiers receive a 20% discount on on-demand Anlas purchases, so heavy anime image generation stays predictable in cost even past the included allowance. A pay-as-you-go option exists for users who do not want a monthly subscription.
Two mechanics around Anlas deserve attention. First, the subscription Anlas pool is a refill, not an accrual: it tops up to your tier's amount each month and only if your balance is below that amount, so unused Anlas never stack. Second, and newer, following the V5 release Anlatan announced that 31 days after launch any unused subscription Anlas will reset to zero the moment a subscription ends — and if you cancel early, they reset on the date the subscription would originally have ended. The company also notes that it was never able to restore a previous month's Anlas when processing a refund.
On refunds, the terms are direct: subscription-based services do not offer refunds unless explicitly granted by the company. Cancelling stops future charges and access continues until the paid period expires. Note also that you cannot purchase additional Anlas once you have cancelled.
Midjourney delivers broader stylistic range and stronger photorealistic and painterly output, but offers nothing on the text-generation side and considerably less granular tag-level control over character attributes. It is the better choice when anime is one style among many you need.
Stable Diffusion running locally, via interfaces such as ComfyUI or Automatic1111, offers complete control, no per-image cost after hardware, and access to a vast ecosystem of community anime checkpoints and LoRAs. The trade-offs are a capable GPU, meaningful setup effort and ongoing maintenance. NovelAI's own image models began as a Stable Diffusion derivative before the V4 generation moved to originally trained models, so the lineage is directly relevant here.
Sudowrite and similar dedicated fiction assistants compete on the writing half with more structured story-development tooling — outlining, character workshops, revision passes — but do not generate images and generally rely on third-party frontier models rather than purpose-trained storytelling finetunes.
General-purpose assistants such as ChatGPT or Claude will out-reason NovelAI's models on analysis, research and structural feedback, and are far cheaper for occasional use. They are more restrictive about subject matter, do not offer comparable long-form fiction finetuning, and their image capabilities are not anime-specialised.
AI Dungeon is the closest comparison for interactive fiction specifically and has a native mobile app, which NovelAI lacks. NovelAI generally holds an advantage in prose quality and context length, and includes image generation that AI Dungeon does not match.
The honest summary: NovelAI wins on the specific combination of anime-tuned image control, long-context storytelling finetunes, private generation and a single flat subscription. If you need only one of those halves, a specialist alternative will likely serve you better and often more cheaply.
This is the most important recent change and it arrived the day after V5 launched. Anlatan announced that Opus subscribers now face usage limits on NovelAI Diffusion V5 specifically, while all other image models and all text models remain unlimited. The stated cause is compute constraint from V5's architecture being 2.5 times the size of V4.5, compounded by what the company describes as a small number of users consuming a hugely disproportionate share of compute.
The limit works as a continuously recharging allowance: every V5 generation at normal resolutions and up to 28 steps draws from it, and it refills hour by hour whether or not you are generating. A full recharge from empty takes about a week, and the company states the capacity is sized so most Opus users will never exhaust it. If you do run out, you can continue by spending subscription or paid Anlas. Both resolution and step count affect consumption, and lowering resolution saves more than lowering steps.
Users should weigh this honestly: the Opus proposition was previously understood as unlimited generation subject to four simple conditions, and on the newest model it is now a metered allowance. The company chose this over raising prices or adding artificial delays, and disclosed it directly, but it is a material change to what the top tier buys.
NovelAI's image models were built on anime imagery whose provenance has been publicly disputed. The Japanese technology outlet GIGAZINE reported in October 2022 that the dataset used for training included many images from the illustration board Danbooru, alongside a technical analysis of NovelAI's modifications to Stable Diffusion — improving CLIP for tag-based prompts, using random cropping instead of central cropping, and tripling the maximum prompt length.
The strongest evidence on this point is not an outside inference but the company's own words. The Taiwanese outlet United Daily News reported in October 2022 that NovelAI's official Twitter account had volunteered the training source in a reply to users: "Since we are training on Danbooru, it also learns character names and their visuals," demonstrating the point by generating a recognisable character from a named anime series. The same report documented an observable consequence — because the training set contained many images carrying artist signatures, generated images began producing signature-like marks of their own, with examples posted by Western and Japanese users in late September and early October 2022. The English Wikipedia entry on Danbooru independently records that NovelAI Diffusion "is reported to be trained on a Danbooru based dataset," and notes the board holds over 10 million user-submitted images as of December 2025.
Danbooru itself has stated its position unambiguously: "Danbooru is not affiliated with NovelAI or any other AI generators, and we don't support or endorse what they do," adding that it is not an AI site and does not use uploaded works for AI training. The site further notes that deleting work from Danbooru will not remove it from an already-trained model, and directs artists to contact NovelAI directly. The English Wikipedia entry on NovelAI records that the service generated considerable controversy over produced imagery, and that Pixiv responded by implementing AI-labelling requirements while declining to eliminate AI-assisted work entirely.
Anyone planning commercial use should factor this history in. It is unresolved rather than settled, and the platform's own terms push the responsibility to you.
The FAQ is candid: NovelAI does not claim ownership of what you create, but "it is possible for the AI to generate content that could be considered copyrighted material," and if you intend to share publicly or commercialise, "it is your own responsibility to ensure it does not infringe on any copyrighted material." The terms reinforce that you retain all rights and ownership of your content while Anlatan waives responsibility for usage outside the service. There is no indemnity offer of the kind some enterprise-oriented image vendors now provide.
Session-based generation means a browser refresh or navigating away from the image page empties your history queue permanently. There is no server-side recovery because there is no server-side copy. For stories, content is encrypted with a key derived from your password and email, and the terms state plainly that a lost password means content cannot be recovered. After a subscription ends, there is a 45-day export window before Anlatan reserves the right to delete stored content.
Prospective users should understand the platform's actual posture rather than inferring one. The terms of service require users to be 18 years or older and capable of entering a legal contract in the United States, but describe no verification mechanism beyond email confirmation. The misconduct clause enumerates prohibited behaviours — reselling, account sharing, circumventing access restrictions, bot abuse, substituting for licensed professions, circumventing regulated businesses such as gambling — and contains no restriction on generated subject matter. The Undesired Content presets are quality filters, not safety filters, and can be disabled outright. The documentation's own framing of the Curated model, recommended for anyone who wants to "make extra sure you don't generate any sensitive images on accident," implies what the Full models can produce.
There is also no NSFW toggle to find. A full-text search of the official image documentation — Basics, Tags, Quality Tags, Undesired Content and UI Functionalities — turns up no occurrence of NSFW, explicit, adult, censor or content filter anywhere. The content boundary is not a switch at all: it is the model you pick (Full versus Curated) and whether you keep the quality presets. Anlatan's own Medium biography states the position more plainly than the product pages do, describing the service as a way to "construct unique stories, thrilling tales, seductive romances, or just fool around. Anything goes!"
The only explicit content prohibition applies to NovelAI Explore, the opt-in public gallery, where uploads face manual review and the guidelines forbid NSFW material, copyright/character/artist tags and advertising. It is worth being precise about what that means: the Explore rules govern public sharing, not what the models will generate privately. Readers should not infer from "No NSFW" in the gallery guidelines that the product declines to produce such material — the restriction applies to the one surface where other people would see it. Anlatan's own FAQ describes the service as offering "an unprecedented level of freedom to the user" and "a limitless playground for your imagination."
One further fact belongs here because of what it implies. There is no official NovelAI application in the Apple App Store: searching for Anlatan returns five unrelated third-party novel-reading apps, and direct lookups on the plausible bundle identifiers net.novelai.app, com.anlatan.novelai and net.novelai all return zero results. The service is delivered purely as a progressive web app, which means no app store has ever reviewed or assigned it an age rating. The 18+ threshold rests entirely on self-declaration. Institutional users, educators and anyone deploying this in a shared or supervised environment should plan accordingly.
At the time of writing, the official image model documentation still lists V4.5 as "Current" while both the homepage and the official blog have announced V5. Separately, the FAQ states text context limits of 3,072 tokens for the lowest tier and 8,192 for the highest, while the subscription documentation and model pages describe 28,672 tokens for Opus. These are contradictions between current official sources rather than an outdated single page, and users should treat the subscription and model pages as authoritative on context size.
There is no documented public API for third-party integration. The platform states it does not currently plan to let users create their own image modules. There are no collaboration, shared-workspace or administrative features — it is a single-user creative tool. And the T5 tokenizer limitation means non-Latin characters, including Japanese, are not reliably supported inside prompts on V4 and V4.5.
There is no substantial body of verifiable independent evaluation. Attempts to retrieve ratings from G2 and Product Hunt were blocked, and the Toolify directory listing shows zero user reviews. A great many "NovelAI review 2026" articles circulate, but the ones surfacing in search are published by competing image-generation platforms and affiliate marketing sites, several of which contain factual errors about the current model generation. None of them are cited here, and readers should discount them accordingly. Third-party traffic estimation from Toolify puts the site at roughly 4.2 million monthly visits with an average visit duration over ten minutes — an estimate, not a disclosed figure, though the engagement depth is consistent with a tool used in long working sessions.
There is a free trial but not a free tier. An email-verified account gets 30 image generations up to 1024x1024, 50 text generations and 100 text-to-speech generations, all one-time allowances that do not refresh monthly. Settings and stories created during the trial carry over if you subscribe. Continued use requires a paid plan starting at $10 a month, or the pay-as-you-go option.
NovelAI does not claim ownership of your content and the terms state you retain all rights to it. However, the FAQ is explicit that the AI can produce material that might be considered copyrighted, and that reviewing output before public or commercial use is your responsibility. The platform waives responsibility for usage outside the service and offers no indemnification. Given the documented disputes over training data provenance, commercial users should form their own view of the risk.
Anlas is the currency consumed by image generation. Tablet and Scroll include 1,000 per month and Opus includes 10,000, refilled monthly to the tier amount only if your balance is below it — they do not accumulate. Opus users can generate without spending Anlas provided they generate one image at a time, use no base image, generate at Normal size or larger, and keep steps at 28 or fewer. All tiers get 20% off additional Anlas purchases. Since the V5 release, unused subscription Anlas reset to zero when a subscription ends.
Yes, according to the official FAQ. Image generations are session-based and not stored on NovelAI's servers, and requests, uploads, edits and generations are not logged. Requests are encrypted in processing. The practical consequence is that you must download anything you want to keep — refreshing or leaving the image generation page clears the history queue with no way to recover it.
Six models are selectable across the V5, V4.5 and V4 generations, each in Full and Curated variants. Curated is trained on a smaller, cleaner, more focused set and is the recommended general-purpose and streaming-safe choice. Full uses a broader, less curated dataset and is the fallback when Curated cannot produce a specific concept. V5 is the newest and most capable, but is the only model subject to Opus usage limits.
Not an official native one. NovelAI runs in any modern mobile browser and can be installed as a Progressive Web App from the site. Several App Store listings use very similar names but are published by unrelated third-party developers and are not affiliated with Anlatan.
Xialong, the current flagship, offers 28,672 tokens on the website plus an 8,192-token rollover window, so usable context in the editor moves between 28,672 and 36,864 tokens as a story grows. Context size is tier-dependent, with Opus receiving the largest. Note that the FAQ page still quotes older, much smaller figures, which contradicts the subscription and model documentation; the latter reflects current models.
Users must be 18 or older per the terms of service, which describe no verification beyond email confirmation — and since there is no official app store listing, no platform has ever reviewed or assigned an age rating either. The misconduct clause covers behaviours such as reselling and account sharing but places no restriction on generated subject matter. There is no NSFW toggle in the interface; the Undesired Content presets filter for image quality, not content, and can be switched off entirely. The one explicit content rule — no NSFW — applies to uploads to the opt-in public Explore gallery, which are manually reviewed. It does not describe what the models will generate privately.
The terms state that subscription-based services do not offer refunds unless explicitly granted by the company. Cancelling prevents future charges and access remains until the paid period ends. Note that you cannot purchase additional Anlas after cancelling, and unused subscription Anlas now reset to zero when the subscription ends.
V5 launched on 21 August 2026 in Curated and Full variants, bringing sharper natural language understanding, expanded multi-language prompting, reworked character positioning for up to 22 characters, higher fidelity, true alpha transparency output and a new Max enhancement level for upscaling. It also brought the first usage limits on the Opus tier, applying to V5 only, because the model's architecture is 2.5 times the size of V4.5.
