Read the four balances before buying more images
The important distinction is percentage allowance versus image credits. An empty V5 usage bar does not mean you must buy a boost: NovelAI says available Subscription or Paid Anlas can still pay for images. Check the charge shown for your selected job. Official usage FAQ.
- Percentage · rechargesRegular V5 usage
- Opus allowance for qualifying Anlas-free V5 jobs. It refills continuously; renewing the subscription does not instantly fill it. Usage rules.
- Percentage · activated boostExtra V5 usage
- A boost adds 100% Extra Usage. Regular usage is spent first and keeps recharging while Extra Usage is available. October 13 announcement.
- Credits · subscription benefitSubscription Anlas
- Refill to the tier allowance at renewal, rather than stacking unused grants. Unspent credits expire when the subscription ends. Refill rules; expiry policy.
- Credits · separate purchasePaid Anlas
- Do not expire. When a job spends Anlas, Subscription Anlas are used before Paid Anlas. Buying these credits is different from buying a usage boost. Anlas guide.
Opus keeps unlimited qualifying images on V4.5 and earlier. V5 instead uses the rechargeable allowance for qualifying Anlas-free jobs. The published size and step limits are a single image, without a base image, up to 1024 × 1024 and 28 steps; larger images or higher step counts already cost Anlas. Eligibility footnote and model-specific rules.
A boost buys a burst of V5 usage, not a new plan
The October 13 launch lists an Opus Usage Boost at $5.99 USD before applicable taxes. You can hold up to four unused boosts; purchased boosts expire after one year. Extra Usage is capped at 400%, and activation requires room for the full 100% addition. Boost announcement; currency and tax terms, §4.1.1.
I would check remaining Anlas and the regular recharge first. A boost may suit a concentrated V5 session. It is less compelling if you can wait, already have credits that will expire, or need settings that make the job charge Anlas anyway. There is no fixed image count per boost: usage depends on the job’s settings.
The September 22 policy announcement converted the existing Subscription Anlas balance to non-expiring Paid Anlas. It did not create a monthly conversion benefit. Canceling early preserves current access and Subscription Anlas until the original paid period ends.
Medium effort saves usage by giving up some controls
The October 8 update adds Medium effort to NovelAI Diffusion V5 Full and its inpainting mode. NovelAI says it uses about 42% less usage at default settings than High at 23 steps. That is the vendor’s comparison; its internal blind tests do not establish equal quality for your prompts. Release and comparison baseline.
Lower usage, fixed settings
- 14 steps and Euler Ancestral only.
- No custom Undesired Content prompt or preset changes.
- Prompt Guidance is adjustable; Prompt Guidance Rescale is unavailable.
The existing full control set
Choose High if your workflow needs a different step count or sampler, a custom Undesired Content prompt, or Prompt Guidance Rescale.
Medium’s savings are not a reason to remove a control the image actually needs.
The toggle sits beside Anime/Furry in the official Effort Toggle guide. The restrictions above apply to Medium, not to every V5 workflow.
Suggested check: keep the prompt and resolution fixed, try Medium, then compare with High at 23 steps. Record the displayed usage cost and whether each result meets your brief. A few usable outputs would support that brief, not a universal quality claim.
For writing, justify the upgrade before paying for images
Paid plans include unlimited text and text-to-speech during the active subscription. The trial is limited: an email-verified account currently gets 50 text and 30 image generations up to 1024 × 1024. Use it to judge whether you like editing continuations. Current subscription guide.
Tablet $10
1,000 Subscription Anlas. My starting paid option if you mainly want prose and do not need a higher-tier model or more context.
Scroll $15
Still 1,000 Subscription Anlas. Consider it for the documented larger text context, not as an image-credit upgrade.
Opus $25
10,000 Subscription Anlas, Xialong access and qualifying Anlas-free image generation. Check the V5 limits above before calling it unlimited.
Published USD prices per renewable 30-day paid period, checked October 16: tier entitlements. Prices exclude applicable tax under the payment terms. Confirm the billed total before purchase.
I would choose NovelAI for an editable story workflow, rather than expect a ready-made companion relationship. You manage prose, lore and generation settings. If character-led chat is the main task, compare the workflow in our SpicyChat review before buying a writing subscription.
A missing story fact may be a Lorebook problem, not a plan problem
The Lorebook can insert stored facts when activation keys appear in recent context, or through Always On. The entry title is not visible to the model. Put the character’s name in the entry text itself. Official entry and activation rules.
For example, an entry titled Mira should contain a complete fact such as Mira is an adult cartographer who keeps a brass compass in her coat. A bare keeps a brass compass leaves the subject unstated.
Proposed diagnostic: check what the model actually receives
- Use a disposable story and enable one short Lorebook entry. Set
Miraas its activation key, then mention Mira in the recent scene. - Open Advanced → Current Context before generating. After a continuation, inspect Last Context for the context used.
- Find the entry’s Inclusion, Reason, Key and Tokens. Turn on Show non-activated entries if the enabled entry is missing from the list.
- If it was excluded, fix activation or context allocation first. If it was included but the output contradicts it, the problem is not simply that the model never received the fact.
The Context Viewer documentation supplies these controls. This diagnostic is suggested, not a test we ran. Inclusion establishes what was sent; it does not guarantee recall, obedience or reliable continuity over a long story.
Check those controls before upgrading for more context. A larger window cannot help an entry that never activates, and filling it with irrelevant lore can make the intended fact harder to manage.
Keep a full story backup; share the file you intend
A .story export retains branches, undo history and retries. Rejected text can therefore travel with it. As Plaintext exports only raw story text; it is useful for sharing the current prose, but is not a full editable backup. For a bulk story archive, the documented path is User Settings → Account → Download All Stories. Export guide.
For image references, do not rely on the FAQ’s blanket “not stored” wording. NovelAI’s current terms, §6.10 and image-caching explanation describe temporary encrypted caching of uploaded inputs. That is the vendor’s account of its processing, not an independent privacy audit.