NovelAI is a subscription AI writing service from Anlatan Inc. It runs its own fine-tuned story models, encrypts your saved stories on your device, and offers a text API with OpenAI-compatible endpoints. We list 3 models, all fine-tuned. They’re unfiltered by policy (NovelAI’s own wording is “No restrictions”), not by abliteration.
This review is based on NovelAI’s published plans, model list, and policies, checked on October 11, 2026. Benchmark results from our own runs will be added when they’re complete. Live details are on the NovelAI provider page.
The short version
| Best for | Long-form fiction and interactive stories, for a fixed monthly price with no per-token meter |
| Uncensored models | 3 (all fine-tuned) |
| Per-token price | None: flat subscription only |
| Plans | Free trial (50 generations), Tablet $10/mo, Scroll $15/mo, Opus $25/mo |
| Privacy score | 33 / 100, joint 22nd of 37 |
| Watch out for | The two large models are Opus-only ($25/mo), and the privacy policy doesn’t address training |
Models
NovelAI’s models are built for prose, not general chat:
- Xialong: a fine-tune of GLM-4.6 (355B parameters) with a 28,672-token context window and an MIT license. It’s the largest model here and has the longest context.
- Llama 3 Erato: a 70B fine-tune of the Llama 3 70B base model.
- Kayra: NovelAI’s in-house 13B model, with an 8,192-token context window.
All three are NovelAI’s own fine-tunes. There’s no abliteration involved: the models are trained for fiction, and the service’s policy is not to filter what you write. NovelAI’s subscription page also names a fourth model, Krake, as Opus-only, which isn’t in our list yet. Search these on the Models page.
Pricing
NovelAI has no per-token prices, so there’s no cost-per-chat figure to quote. You pay a flat monthly fee, and every paid tier includes unlimited text generation. What changes between tiers is which models you get and how much of your story the model can see at once (its context).
- Paper (free trial): 50 text generations with an email-verified account. Your stories and settings carry over if you subscribe.
- Tablet, $10/mo: unlimited text on the standard models. Xialong and Erato aren’t included, so from our list that means Kayra.
- Scroll, $15/mo: the same models as Tablet with a larger context. NovelAI doesn’t state the exact size on its subscription page.
- Opus, $25/mo: unlocks Xialong and Erato, with context up to 28,672 tokens.
The plans also include “Anlas”, NovelAI’s currency for image generation: 1,000 a month on Tablet and Scroll, 10,000 on Opus. If you only want text, you can ignore Anlas.
For most writers the real choice is Tablet or Opus. Tablet is the cheapest way in. Opus is the only tier with NovelAI’s two large models, and it’s the one most people comparing it with API hosts will want.
The text API is useful for front-ends, but its terms say each generation must be started by a human action. Bulk or automated generation breaks the terms, so it’s not a substitute for a pay-per-token API.
Privacy
NovelAI scores 33 / 100 on our privacy rubric, which only gives points for documented claims:
| Rubric item | NovelAI | Points |
|---|---|---|
| Prompt retention | Says it doesn’t store prompts; technical logs kept up to 90 days | 10 / 20 |
| No training on prompts | Not addressed in the policy | 0 / 15 |
| Technical guarantee (TEE / E2EE) | Saved stories are encrypted on your device; generation relies on policy | 0 / 20 |
| Signup | Account required | 3 / 8 |
| Crypto payments | Card only | 0 / 7 |
| Jurisdiction | Massachusetts, US | 10 / 10 |
| Upstream hosts disclosed | Yes | 10 / 10 |
| Independent audit | No | 0 / 10 |
That’s joint 22nd of 37. The strongest point is storage: stories are encrypted client-side with a key derived from your password. That protects what you keep, not what you send. The text you submit for generation is still handled under the privacy policy, and that policy is silent on training. The 90-day technical logs are why retention scores “minimal” rather than “none”.
Access and payment
- Web app for writing, plus a text API with OpenAI-compatible endpoints.
- Account required.
- Card only. We found no crypto option.
- Run by Anlatan Inc., in Massachusetts, US.
- API use must be human-initiated per its terms.
How it compares
- Flat rate with a real API: Infermatic is $9/mo for API access to curated roleplay fine-tunes up to 72B, takes crypto, and scores 40 on privacy. Arli AI starts at $10/mo and we list 131 models there.
- Cheaper entry: DreamGen starts at $7.83/mo, but it’s credit-based rather than unlimited, with a 5,000-token context on that plan.
- Pay per token for the same kind of model: AionLabs sells its own storytelling models from $0.70 / $1.40 per 1M tokens (input / output) and scores 48 on privacy.
Verdict
NovelAI is a good fit if you write fiction in its editor and want predictable costs. Unlimited text from $10 is simple to budget, the models are its own, and encrypted story storage is a genuine plus. Opus at $25 is the plan to pick if you want Xialong’s 355B parameters and 28K context.
Look elsewhere if you need a general-purpose API, crypto payment, or automation. The API terms rule out bulk use, and the privacy policy doesn’t say whether your prompts can be used for training.
See NovelAI’s models and plan details on its provider page, or compare every provider in the index.