Nouveautés d’Underfiction
• Un tour continue de s'écrire si la connexion tombe ou si tu quittes l'application, et reprend là où il s'était arrêté à ton retour. • Arrêter un tour aboutit même quand le réseau avale la première tentative. • Le journal indique ce que coûtent les images, les clips et la narration, pas seulement les tours. • Les histoires ne se renomment plus en plein chapitre, et une histoire s'ouvre sur son tour le plus récent. • Corrections un peu partout.
• Corrections de bugs et améliorations de la stabilité.
Gemini 3.8 Flash takes the Flash slot and stays the default story model — the same 1M context and the same rate per token as 3.7. It does think longer before it writes: measured on the same scenes, deep in a long chapter it spends about twice the hidden reasoning 3.7 spent for the same paragraphs, so the picker prices Flash at three credits per thousand characters instead of two. Turns already written on 3.7 keep its name in the log.
MiniMax H3 Max Turbo joins the models you can animate on: the same five to fifteen seconds, the same 480p and 768p, sound on every clip, at half of H3 Max's price. Five seconds at 768p costs twelve credits, against twenty-four on H3 Max and a hundred and sixty-four on Grok. H3 Max keeps its place in the picker.
• GLM 5.2 et Grok 4.6 rejoignent les modèles recommandés : le tour le moins cher, et le moins de refus. • MiniMax H3 Max anime des clips avec son pour un tiers des crédits.
MiniMax H3 Max joins the models you can animate on: five to fifteen seconds, 480p or 768p, sound on every clip, and well under half the price of anything else here that carries sound. HappyHorse leaves the picker in the same breath — MiniMax does its job for a third of the credits. Clips already animated on it keep its name.
• Publie histoires, personnages et mondes sur l'étagère de la communauté — et parcours ce que d'autres y ont posé. • Les guillemets font d'une ligne une réplique, la barre oblique dirige la scène, la prose nue reste prose. • En Roleplay, leur tête est ouverte ; la tienne reste la tienne. • L'ouverture que tu as tapée survit au passage Story-Roleplay. • Les étagères trient du plus récent. • Sur Android, les portraits quittent le fichier de la bibliothèque : chargement léger, frappe fluide.
• Le champ où tu écris flotte au-dessus de la prose, avec la place d'un paragraphe. • Un prompt d'image arrive en une seconde environ, écrit mot à mot. • Choisis l'image à animer sur une étagère : scène, personnages, monde, tes photos. • Écris une ligne au-dessus de la description et génère le reste. • Une seule sorte de consigne par histoire : écris-la, nomme-la, utilise-la partout. • Grok 4.6 et Gemini 3.7 Flash prennent leurs places ; les images, Grok Imagine 2.
Gemini 3.7 Flash takes the Flash slot and stays the default story model — the same 1M context and the same rate per turn as 3.6, on a model Google built to hold a longer thread. Turns already written on 3.6 keep its name in the log.
Grok 4.6 takes the Grok slot in the model picker, still set to write rather than deliberate. Illustrations get Grok Imagine 2, and Grok clips animate on Grok Imagine 1.5, which adds 1080p. Turns and images already made on the older versions keep their names in the log.
• Histoire ou Jeu de rôle : choisis la forme à la configuration. Le rôle s'ouvre sur un personnage. • Les contrôles de narration nomment leur sujet ; le narrateur à la première personne garde son intériorité. • Écris une consigne une fois, utilise-la partout. Les consignes enregistrées se synchronisent. • Ta bibliothèque te suit sur chaque appareil où tu te connectes. Un bouton coupe la synchronisation. • Nomme ton protagoniste avant que l'ouverture s'écrive.
Grok 4.5 rejoins the model lineup, set to write rather than deliberate — it reaches the first line of prose in seconds. Kimi K3 leaves the picker; turns already written on it keep their name in the log.
• Cinq classiques rejoignent la bibliothèque : Orgueil et Préjugés, Jane Eyre, Dracula, Persuasion, Sherlock Holmes. • La bibliothèque s'ouvre sur une vitrine de mondes et d'ouvertures, repensée pour mobile. • Choisissez la narration, ou une consigne permanente que chaque tour suit. • Joignez une image : le narrateur écrit d'après ce qu'il voit. • Tout passage peut être lu à voix haute, dans la voix choisie. • Exportez une histoire ou la bibliothèque entière en .underfiction portable.
The Library has its own shape on a phone now: a showcase that fits the width, a menu that sizes to its labels, and detail views with one clear way back. The navigation menu stays put on the pages it opens.
Credit amounts in logs, generation details, and top-ups are written in the language you are reading, on web, iOS, and Android.
A character card or archive that will not open now names the file and the reason, instead of failing with one generic line.
The Android model picker shows what each model costs per turn, matching web and iOS.
The starter browser is native now: page through the built-in worlds, their leads, and every authored opening without leaving the Library.
Narration nudges are called directions, and story setup groups them into sections instead of one long list.
Copying a character asks first and leaves the editor open, instead of dropping you back to the list.
Choose the voice before you write: First Person, Second Person, Present Tense, Inner Thoughts. The narrator holds what you pick for the rest of the scene, and scene summaries now keep that voice instead of pulling the prose back toward third person.
Give a story a direction in your own words — "keep her unreliable", "no dialogue tags" — and every turn follows it. It outlasts compression, travels inside portable archives, and comes along when someone takes the pen on a shared excerpt.
Pip, Syra, Lydia, and Lola each get an authored opening built on a shared secret or a craft rather than a stated romance. They lead their worlds in the library rail, and The Antler & Stone reaches the showcase for the first time.
Story and narration model pickers mark which models run inside Venice and which reach the model through anonymized routing. Hover a badge and it spells out what that guarantee covers.
Stories, characters, and worlds are held locally, and nothing about them leaves the device unless you move it yourself. Portable .underfiction archives are how a story or a whole library travels between web, iOS, and Android.
The prose composer uses a single compact layout from phone to desktop, and long speaker names no longer crowd out the field.
Fixed iOS navigation dismissal and zoom transitions, web header and scrollbar shifts, Android segmented-control sizing, and a setup layout that moved while an animation ran.
The starter grids are now a moving showcase on the landing page and in every Library. Lead characters frame each world, the openings read like a table of contents, and the stage advances on its own until you touch it.
Pick a world or a character and you land in story setup. The onboarding walkthrough and the separate persona step are gone.
Export everything — stories, characters, worlds, images, videos — as one portable .underfiction file, and import it on any other device. Merging keeps whichever copy of a story you edited last.
If you have not written anything yet, your shelf opens on the built-in worlds and characters instead of a blank screen.
The complete classics collection — worlds, canon leads, and openings — is now localized in all eight supported languages.
Rebuilt around the showcase, with starting in the browser as the first move and the native apps a quieter link below it.
Pride and Prejudice, Jane Eyre, Dracula, Persuasion, and Sherlock Holmes joined the built-in library on web, iOS, and Android — eight canon leads, new cover art, fifteen authored openings. You are not the protagonist by default: you take the governess's chair in Jane Eyre, Anne's position in Persuasion, and elsewhere you arrive beside canon as a confidante or a client.
The story model picker opens on a clear first group of recommended models, with ratings and prices lined up down the list. Kimi K3 is back in the lineup.
Any illustration already attached to the story can be the source when you animate a scene, not just the most recent one.
Media actions use the same labels and sit in the same place on web, iOS, and Android.
The prose language follows what you write, independent of the interface language. Switch mid-story and the narrator switches with you; keep going and it stays in the scene's language. The per-story language setting is gone, and interface language moved into Preferences.
Story cards show the artwork of the world they belong to again.
Rerun, edit, share, and the rest stay quiet until you reach the passage they belong to.
Tuned story setup, character editing, settings, navigation, and the input bar so web and iOS read as the same product at the same rhythm.
Attach an image to your next turn and the narrator writes from what it sees — a face, a room, a photograph you want in the scene. On web, iOS, and Android.
Have any passage narrated in the model and voice you choose. Narration is cached on your device, so replaying a passage is free.
Take a story whole, or stopped at any earlier turn, as readable Markdown or as a portable .underfiction archive carrying its characters, world, images, and videos.
Import, export, and the rest of the library actions moved out of the page body and into the toolbar.
Credit spend rounds the way you would read it, and model pricing labels in story setup say what they mean.
Fixed menu labels that fell back to English and a language switch that did not always take.
Upgraded Claude Opus from 4.8 to Opus 5 — a step up in long-scene coherence and prose control at the same price. Context window and credits-per-turn are unchanged; existing stories pick it up automatically.
New starter stories on the landing page and in Library — curated openings you can step into with one tap, now on web, iOS, and Android. The library gained a Mine/Library switch so built-in material and your own work stay cleanly separated. Gemini 3.6 Flash joined the model picker and became the default story model.
• Relancez un tour et vos versions alternatives sont conservées : elles survivent à une annulation, un échec ou un retour en arrière. • Les longues scènes coûtent moins cher à poursuivre et tiennent mieux, grâce à une compression plus intelligente en arrière-plan. • Kimi fait son retour. • Et une série de correctifs de fiabilité pour l'enregistrement et la génération.
Share a moment from any story as an unlisted link — a directive and the prose it produced, in a clean reader view. Readers can take the pen: one tap copies the full setup into their own library so they can continue the scene their way. Links are unlisted and never indexed; only you can share them.
• Les histoires qui partent d'une scène vierge écrivent désormais leur ouverture au lieu d'échouer. • Réessayer après une génération échouée ne duplique ni ne perd plus de passages. • Quand un modèle refuse une scène, vous voyez désormais sa vraie raison au lieu d'une erreur générique. • Un enregistrement plus léger et plus fiable, avec moins de messages d'erreur.
• Les livres de personnage agissent désormais dans l'histoire : leurs entrées se déclenchent sur leurs mots-clés au fil de la scène. • L'import de cartes Tavern est maintenant sans perte — scénario, salutations et livres survivent à l'aller-retour. • Fiches de personnage réordonnées selon ce qui guide vraiment la prose, avec une nouvelle section de détails. • Messages plus clairs quand une génération échoue ; les générations vides sont remboursées automatiquement. • Personnages allégés dans les mondes de départ, corrections du flux de crédits et la tournée habituelle de finitions.
• Illustrer : générez des images de vos personnages et de vos scènes, directement depuis l'histoire. • Animer : transformez une illustration ou un arrière-plan en court clip vidéo. • Nouveaux modèles : GPT 5.6 Terra, Kimi et DeepSeek — Seedream 5 Pro prend la tête de la génération d'images. • Un composeur plus net : réglages, Illustrer et Animer regroupés derrière un menu « + ». • Revenez au bas de la scène d'un geste, avec une bibliothèque plus rapide et les correctifs habituels.
• Fonds de scène : posez l'ambiance avec une image de fond, activable depuis la barre de saisie. • Nouveau navigateur de scènes : passez d'un chapitre à l'autre depuis le titre de l'histoire. • Importez des cartes de personnage Tavern (V2 & V3) et des world books sur toutes les plateformes. • Bibliothèque épurée, réglages simplifiés, traitement des images plus rapide — et les corrections habituelles.
• Nouveaux modes d'ouverture: Adapter, Littéral et Personnage. Les personnages peuvent maintenant enregistrer des salutations et les importer/exporter via TavernCard first_mes.
• Le lot habituel de corrections de bugs et de finitions.
• Underfiction parle maintenant japonais, coréen et chinois traditionnel. • Et le lot habituel de corrections et de finitions.
• Underfiction parle maintenant espagnol. • Et le lot habituel de corrections et de finitions.
Upgraded Opus from 4.6 to 4.7. Bumped the freedom rating on GPT and the quality rating on Gemini to reflect recent provider improvements. Fixed a credit-purchase webhook bug where some Stripe deliveries silently retried instead of posting credits cleanly on the first try.
Overhauled the credits experience with cleaner controls, a smarter default amount, and instant rendering of saved cards and auto top-up. Redesigned the landing page with a full-screen hero, prose sections, and a simplified world showcase. Fixed missing tab bars when navigating directly to /characters or /stories, and resolved a handful of hydration errors. Loading skeletons now match the actual content they replace.
Removed the mood system entirely — presets, prompts, and UI — to simplify story setup and let your prompts shape tone directly. Stabilized paragraph reveal during streaming so text lands smoothly. Improved the main menu, fixed a hero image flash on world open, an edit-last-turn bug, and assorted mobile polish issues.
Normalized spacing across galleries, modals, and page shells, and tightened touch targets on mobile. Switched to new generative cover art across the worlds.
Introduced shared typography, control, surface, and spacing token systems across the app for consistent rhythm and visual language. Refreshed the logo, favicons, and promo video, and polished the navigation icons.
Built-in worlds and characters are now immutable templates — editing one transparently clones it to your personal library on first edit, so the originals stay clean. Added a dedicated My Worlds section, made the Core sections collapsible, and refined character card defaults. Added a Sentry alert for missing Venice pricing so model gaps get caught before users hit them.
New fast-track flow lets you start a story in seconds — pick a world, drop in a character or persona, and you're writing. Refined the mobile bottom navigation, polished modal interactions, and added new crosspost articles to the blog.
Image previews now appear instantly, with no shimmer flicker when changing sources. Embedded a promo video on the landing page and refreshed the model picker.
Export a world and all its characters as a portable book file, and import books from other users with their images included. Added comprehensive SEO infrastructure with structured data, blog posts, homepage FAQ, and an accessibility pass that brought the PageSpeed score to 100.
Stabilized text reveal streaming so paragraphs land without jumping. Fixed shimmer flicker on cached character images, resolved leaked body scroll locks, and improved overall scroll performance on long stories.
Rewrote scene summary prompts to better preserve character dynamics and relationship arcs across transitions. Unified the scene opening flow so new scenes, forks, and regenerations all start consistently. Added user-friendly provider error messages that name the model and offer a one-tap retry.
Rebuilt the core story runtime for reliability and lower token cost. Menu pages now open as overlays inside the story view so you never lose your place. Added mid-scene history compression to keep long conversations affordable, and introduced a free-tier framing for new users.
Fixed race conditions in the billing pipeline and library reconciliation. Patched a cross-user data leak, closed an auth bypass on system-generated content, and added missing database indexes. Improved cursor stability during streaming.
Introduced Patronage loyalty tiers with bonus credits on every purchase, and a configurable auto top-up so you never run dry mid-scene.
Import and export characters as TavernCard V2 PNG files — fully compatible with SillyTavern, ChubAI, and other platforms. Added smooth expand/collapse on the advanced character fields.
New bottom navigation bar with thematic icons. Scene list and menus now open as swipe-to-dismiss bottom sheets. Replaced the input modal with an inline expanding input bar. Scene suggestions are now free for everyone. Fixed a handful of tap-target and zoom issues on iOS.
Launched the referral program: invite a friend and you both get 1,000 credits when they make their first purchase.
Restructured the character model as a TavernCard V2 superset — personality PLists, example dialogue, post-history instructions, and full compatibility with the broader roleplay ecosystem. Replaced dollar-based billing with credits, a purpose-built currency with clearer pricing.
Worlds and characters now load from markdown content files — a single source of truth that's easy to author and version. Redesigned the landing page around worlds, characters, and creative freedom. Added per-world hero images and a mood system that shapes the AI's tone.
Dedicated Characters tab with portrait cards organized by world. Full-bleed character portraits in the edit modal. Separated built-in characters from your personal library so edits to one don't affect the other. Fixed hero image flicker on first world open.
Introducing Worlds — curated world settings with their own characters, lore, and narrative hooks. Four starter worlds ship out of the box: Ashworth End, Linden Park, Antler Lodge, and Marabou Syra. Each world generates dynamic scene suggestions via the LLM so every playthrough starts differently.
The platform has a new name and identity. What started as Fablis, then became Unreliable, is now Underfiction — a name that better captures what we're building. Added Sentry error tracking with full content redaction, and a pre-flight context window check to catch token overflows before they time out.
Published three launch-day posts covering why we built this, the case for local-first privacy, and what literary AI prose should actually read like. Enabled anonymous free tries so new visitors can start a story before signing up.
Generation now falls back to Grok when Venice hits rate limits, so you're never stuck waiting. Defaulted everyone to Opus with an 8k response length. Added loading skeletons, API caching, and entrance animations across the app.
New public landing page that showcases the platform to visitors. Unified all page headers into a shared PageHeader component and reworked logo routing. Replaced anonymous usage stats with per-user credit and usage data.
Underfiction is live. Five preset story scenarios to get new users started, a welcome credits bonus on signup, and the password gate is gone. Extensive mobile polish: 44px tap targets, proper iOS keyboard handling, fixed content overlap with the floating header.
Five curated story presets — mainstream genres with pre-built characters and prose samples — appear on the dashboard for new users. Added an advanced settings section with editable system prompts for full control over the AI's behavior. Fixed a bug where the scene auto-generation never triggered.
Switched to full turn history instead of a truncated prose blob, giving the AI much richer context. Added optional example dialogue and style reference prose to character cards. Two-tier scene summaries: short for the sidebar, detailed for context injection.
Paper texture, upgraded drop caps, and scene ornaments give the reading view a tactile, book-like feel. Card hover animations, a streaming cursor, and prose reveal fades round out the mood. Also fixed the most persistent scroll-jump bugs during text reveal.
Full mobile-responsive pass with touch-optimized interactions and viewport fixes. Added SEO infrastructure: OpenGraph images, JSON-LD, sitemap, and robots.txt. New settings page with model picker, temperature slider, and max response length control.
Choose between roleplay, balanced, and novel prose styles — each shapes how the AI writes and responds. Mixed narration and dialogue input with distinct visual styling. Adaptive narration so the AI matches your writing cadence. Also added an in-memory story cache for instant page transitions.
Fork any conversation from any turn to explore alternate paths. New slash command menu: /as to switch character voice, /tone to shift mood, /newScene, and /narrate. Warmed up the color palette and added a subtle prose surface background.
Pay-per-use credit system powered by Stripe with reserve-then-settle billing. Usage logs page with model breakdown and provider cost. Transaction history with Stripe receipt links. Custom amount top-ups and an auto top-up toggle.
Sign in with Google or GitHub, or create an email/password account. Password reset via email. Automatic OAuth account linking when emails match. Migrated from SQLite to PostgreSQL for production reliability.
Stories can now span multiple scenes with automatic async summarization to keep context manageable. Migrated local storage from localStorage to IndexedDB, which handles much larger content without hitting size limits.
Added a light/dark theme toggle that respects system preference, with smooth 150ms color transitions. Switched the reading typeface to Literata for a warmer, more bookish feel. Fixed dark mode hover states across the UI.
The first working build of what would become Underfiction. Core story engine with AI-powered prose generation via Venice.ai, a reading view, character creation, and scene setup. Stories saved locally in the browser. Everything starts somewhere.