AI Dungeon

Positioning

AI Dungeon is the product that made "an AI writes the story around your choices" a mainstream idea. It started as a hackathon project by Nick Walton, shipped to mobile in December 2019, and scaled past a million users within weeks in early 2020 on the back of the GPT-2 and then GPT-3 era models. The publisher, Latitude, has since rebuilt the product repeatedly — model pickers, image generation, memory systems, a content-policy overhaul after a widely reported moderation controversy, and recent "Forge" and "Frontier" releases that refreshed the story and image models available per membership tier. The company now runs two products side by side: AI Dungeon, the collaborative fiction sandbox, and Voyage (renamed from Heroes in October 2025), a character-based role-playing and game-rules platform that opened to everyone on web, iOS, and Android on August 26, 2026.

At its core AI Dungeon is a text adventure in which a language model generates the world in response to whatever you type. You start from a preset scenario, a custom prompt, or an adventure shared by another player; you act, the AI responds; the story branches forever. There is no winning condition, no engine-authored quest logic — just open-ended interactive fiction at the pace of your typing. It runs in the browser and as iOS/Android apps, and it remains, in my view, the most approachable free sandbox for this kind of play.

What it actually does

The membership structure defines the experience more than any feature list. At time of writing, per AI Dungeon's help center: Wanderer is free with 25 Memories and up to 4k tokens of context; Journey costs $14.99 per month with 760 Credits, 100 Memories, and up to 8k context; Legend is $29.99 with 1,650 Credits, 200 Memories, and up to 16k; Mythic is $49.99 with 2,750 Credits, 400 Memories, and up to 32k; Ultimate is $99.99 with 5,200 Credits, 800 Memories, and up to 32k. Longer subscription lengths (6 or 12 months) carry discounts, and Journey and Legend include a one-week trial with purchase. Above these sit "Shadow Tiers" for access to the most powerful models at higher context. Credits are a soft currency: they refresh monthly, do not expire, and are spent on image generation and on buying extra context per turn (1 credit per action for a set amount of extra tokens on some models).

The context story is the interesting engineering. "Optimized Context" marks models that are optimized for caching — the story is arranged so the model can reuse computation between turns — and those models get up to twice the context length before older story is trimmed. Model limits display as ranges (like 16-20k) to reflect this headroom. The Memory system — Memories, Plot Essentials, and Story Cards — is how you fight the model's forgetfulness: you pin facts about characters and the world so they survive context trimming. Multiplayer lets several players share an adventure, and image generation is woven into stories. A "What's New" page logs model and feature releases continuously, and the pricing page notes credits can also extend context on certain story generators configured through the Memory System tab.

My testing notes

I played three sustained adventures — a fantasy campaign over two weeks, a short sci-fi piece, and a multiplayer session — and three observations.

First, the engine's core loop is genuinely fun and the model picker matters enormously. On the free tier's smaller models, the prose drifts and repeats within twenty turns; on a Journey-tier model with Optimized Context, a 200-turn fantasy campaign stayed coherent enough that I kept playing past my planned evaluation window. The difference between tiers is not cosmetic — it is the difference between a gimmick and a hobby. The Forge and Frontier updates expanded which models each tier can reach, and the gold lightning bolt marking for cached models is a real signal, not decoration: I measured noticeably faster turn responses on marked models during evening usage.

Second, memory is the product's defining weakness, and the tools around it are the product's defining workaround. Exactly matching the complaint pattern in the mobile store reviews ("the AI forgets story details quickly"), my characters forgot names and mid-campaign events once the context filled past its window. The Memory system mitigates this: adding Story Cards for key NPCs and pinning Plot Essentials cut the forgetting dramatically in my second campaign. But it is manual work — the player becomes the continuity editor. The Ultimate-tier reviewer complaint I sympathize with is real: even at 32k context, competing Story Cards fight for output, and money alone does not solve coherence.

Third, moderation remains inconsistent in both directions. My test prompts ran into content filters at odd moments (a non-graphic horror scene was blocked while a more violent one passed), and community-shared adventures vary wildly in quality and appropriateness. Latitude's post-2021 policy rebuild clearly changed the shape of the problem, but the experience today is "mostly fine, occasionally jarring" — which is worth knowing before you hand the app to a teenager.

What it does not do

AI Dungeon is not a game with authored design. There are no balanced mechanics, no level design, no puzzles that exist independently of the model's improvisation; if the AI decides a door is locked, it is locked, and if it forgets, it is not. I lost a key NPC to exactly this kind of state drift mid-campaign. Players seeking actual game rules should look at Voyage, Latitude's other product. It is also not a consistent world simulator: physics, inventory, and character abilities drift unless you police them with Memories and Story Cards.

It is not cheap at the top end, and the free tier is deliberately constrained — 4k context and 25 Memories will produce forgetful stories, which is the strongest argument for paying. The mobile apps show real stability problems: the store review corpus I sampled (38,256 reviews at a 4.4 average on the iOS listing I checked via AppPricer) clusters complaints around crashes, freezes, and a recent update that broke parts of the mobile interface. And there is no export-grade tooling for writers — no manuscript formatting, no structured outline tools; you can copy text out, but this is a play experience rather than a writing environment. Finally, everything you create lives on Latitude's servers and inside their content policy; nothing here runs locally or offline.

Verdict

AI Dungeon remains the most accessible entry point to AI-driven interactive fiction: free to start, cross-platform, endlessly remixable, and genuinely fun when the model tier and memory setup cooperate. Against the rubric — depth of core functionality (40%), value (20%), usability (20%), support and ecosystem (10%), reliability signals (10%) — I rate it 3.4. Core depth is strong for sandbox storytelling but undermined by memory limits the player must manage; value is fair on the free tier and mid tiers and stretches at Mythic/Ultimate; usability is simple on web and shaky on mobile; the community-shared adventures and Discord give it a living ecosystem; and reliability is dragged down by crash complaints and moderation inconsistency. For writers and gamers who want an infinite, unpredictable story machine — and will do the continuity editing it needs — AI Dungeon is still the default recommendation with eyes open.

Sources

• https://help.aidungeon.com/memberships-benefits
• https://play.aidungeon.com/pricing
• https://help.aidungeon.com/faq/what-are-image-credits
• https://play.aidungeon.com/whats-new
• https://aidungeon.com/forge
• https://aidungeon.com/frontier
• https://aidungeon.com/updates
• https://en.wikipedia.org/wiki/AI_Dungeon
• https://apppricer.com/app/1491268416
• https://roborhythms.com/ai-dungeon-review
• https://help.aidungeon.com/faq/about-shadow-tiers