AI Dungeon Review 2026: Still the Best Free Text Adventure?

Ottoline Fennick kept her secret for about forty turns. Her husband, the last harbormaster of Saltmere, the salt-marsh town I'd built in AI Dungeon, hadn't drowned. He'd rowed off with eleven years of customs money, and she'd been covering for him ever since. I typed all of that into her Story Card myself, with one rule underneath: she lies about this until the inspector finds the second ledger.

Then my inspector asked, idly, whether her husband had been a good sailor. Ottoline put down her cup and confessed the whole thing. The theft, the rowboat, the forged entries. In one paragraph. On the third night of a twelve-day run.

It took me a day to work out why, and that explanation is the reason this AI Dungeon review exists. We'll get to it. I first poked at Latitude's text adventure game for a couple of weeks in spring 2025, back when I was new to all of this, and came back over the summer for a fantasy campaign roundup where it finished second. This time I gave it a story to itself: Tuesday, September 22 through Saturday, October 3, 2026, six days on the free tier and six on the $14.99 Journey plan, roughly 330 actions in total.

How a turn works: Do, Say, Story, See

Almost every app I review on this site is a chat window. You type, a character answers. AI Dungeon looks similar for about ten seconds, and then you notice the little mode switcher next to the input box, and that switcher is where the design lives.

Latitude's how-to-play guide describes four input modes, and they really do change what comes back:

• Do makes your character act. “Lift the floorboard” becomes “You lift the floorboard” and the AI narrates the result.
• Say puts your words in your character's mouth, so the people in the scene respond to dialogue instead of narration.
• Story lets you write the narration yourself and has the AI pick up from there. This is the one experienced players lean on.
• See generates an image for the adventure. It usually costs credits, which matters on the free tier (more below).

Below that sit three buttons: Continue (the AI writes another beat without you), Retry (a different version of the last output, stacked so you can flip back), and Erase. I used Retry constantly, maybe one turn in five, and Continue more than I expected, because letting the narrator run for two beats is how you find out what the world wants to do when you're not pushing it.

The split pays off. When my inspector walked into the tide office with Do, the narrator described the room: tide tables pinned to the wall, a kettle that had been boiling too long. When I switched to Say and asked Ottoline about the tables, she answered as a person with an agenda. Same input box, same story, very different writing. Chat apps blur those together, and it's a big part of why their “adventures” drift into conversation. (If you're more writer than player, the Character.AI for writers piece covers the chat-first way of doing story work.)

One complaint. The default mode on a fresh adventure is Do, and the app remembers your last mode per adventure, which sounds sensible until you realize you've sent three lines of dialogue as actions and your inspector has been physically doing the sentence “I'm not here to cause trouble.” Minor, and it still got me four times.

Is AI Dungeon free? Free with no turn cap

The free tier is called Wanderer, and I want to be clear about this because half the search results for “is AI Dungeon free” are years out of date: there's no energy bar, no daily message counter, and no ads. The App Store listing describes it as free-to-play without ads, and in six evenings (about 180 actions) I never ran into a wall.

Here's what Wanderer includes, per Latitude's memberships page and model list as of early October 2026:

• 2K to 4K tokens of context, depending on the model (Muse, Wayfarer Small 2, Fable, and Shimmer get 4K; Harbinger, Madness, and Hearthfire get 2K)
• 25 Memories in the Memory Bank per adventure
• Zero monthly credits
• Nine Basic models: Dynamic Small, Muse, Wayfarer Small 2, Madness, Fable, Shimmer, DeepSeek V4 Flash, Hearthfire, and Harbinger
• Story Cards, Plot Essentials, Author's Note, and AI Instructions

That context number is the real limit, and on the free tier it's small. A context window of 4K tokens is very roughly 3,000 words, and 2K is half that. It has to hold your instructions, your active Story Cards, the summary, and the recent story all at once. By the end of the first free evening Saltmere had five Story Cards and a Plot Essentials block, and I could feel the recent past getting squeezed, worst of all on Harbinger. Details from twenty turns back were gone unless a card held them.

And the zero credits matter for one specific thing: images. The See mode guide says creating images usually costs credits unless your tier or the image model says otherwise. Free players can buy credit packs, but “a free tier with no images” is the fair description.

Compared with the free roleplay chatbots I've tested this year, that's a generous deal. Most of them ration messages or wedge ads between replies. If unlimited messages are your top requirement, AI Dungeon belongs on the same shortlist as the apps in my no-message-limits roundup, even though it isn't a chat app at all.

AI Dungeon models: the free list, and what $14.99 adds

Latitude sorts its AI Dungeon models into three buckets. Basic models run on every tier, free included. Premium models start at Journey. Ultra is a small top shelf, currently DeepSeek variants. The list changes often enough that the help center keeps a “deprecated” section (Hermes 3 405B is in it right now, hidden behind a toggle) and an archive of removed ones like Mistral Large 2.

The part that isn't obvious from the pricing page: context is set per model, not just per tier. The tier numbers you see advertised (“up to 4K,” “up to 8K”) are ceilings, and each model gets its own slice under them.

The free models I actually used

I tried eight of the nine Basic models (all but Madness, more on that below) and settled on two. Harbinger is described as an adventure finetune where choices matter, and it lived up to that. When my inspector bluffed his way past the night watchman, Harbinger had the watchman check the bluff the next morning. Nothing else on the free list did that unprompted. Its catch on Wanderer is that 2K window.

Muse is pitched at character work, and I switched to it for long stretches with Ottoline, partly because her voice stayed drier and more guarded there and partly because Muse gets twice Harbinger's context on the free tier. Wayfarer Small 2 is the combat-and-consequences model; I used it for one tide-office fire scene and it was vivid and a little cruel, which was the point. DeepSeek V4 Flash is the newest addition and good with fast dialogue. Madness is tuned for darker material, and I skipped it entirely, since Saltmere is a customs mystery, not a horror story.

Dynamic Small randomly rotates between several models. I get the idea (it fights repetition), but in a mystery where tone consistency is half the fun, it felt like a different narrator every few turns. Turn it off if you're running anything with a mood.

What changed when I paid

On Monday, September 28 I paid $14.99 for Journey. The Memory Bank went from 25 to 100, a monthly bundle of 760 credits landed, and the Premium list opened up, including Wayfarer Large, Nova, Equinox, Shimmer Pro, Gemma 31B, GLM 5.3 Flash, Dynamic GLM, and DeepSeek V3.1 and V3.2.

Context moved too, unevenly. Harbinger went from 2K to 8K, and that was the change I felt every single night: the bluff, the fire, and Ottoline's lies all stayed in view at once. Muse went to 16K. The big Premium models, though, start at just 4K on Journey. Wayfarer Large, Nova, and the DeepSeek V3 models all get 4K there and only climb at Legend and above.

I still found Wayfarer Large worth switching to for crowded scenes, even at 4K. It kept three side characters distinct in the same room (Ottoline, the watchman, and a fish-seller who hated everyone), where the free models tended to merge any two who were standing close together. For long stretches I went back to Harbinger and its bigger window. DeepSeek V3.2 wrote prettier sentences and was worse at letting my inspector fail.

The 128K figures on the model list are ceilings at the very top tiers, not what a Journey player sees. Credits can extend context temporarily on some models, but every action at extended context costs credits, including retries, and I retry a lot.

My regret, since every month of testing seems to produce one: I spent a chunk of the 760 credits on See-mode portraits of Ottoline on the first paid night. Four attempts. She came out as a different woman each time, once with a lighthouse growing out of her shoulder. I should have saved them for extended context during the finale. The credits didn't come back.

Story Cards, memory, and the two times Saltmere broke

AI Dungeon gives you more memory tools than any of the chat apps on my phone, and they do different jobs. The settings guide lays them out:

• Story Cards: entries for characters, places, and factions that get pulled into context when their trigger keywords show up. Same idea as a lorebook.
• Plot Essentials: the facts you want the AI to see all the time.
• Author's Note: genre, tone, style.
• AI Instructions: rules sent at the very start of the context.
• Memory Bank: automatic snapshot memories of events, recalled when they seem relevant.
• Auto Summarization: writes and keeps updating a Story Summary as you play.

If you've built anything in a lorebook before, Story Cards will feel familiar, and most of the entry patterns in my Character.AI Lorebook ideas post carry straight over. (The memory explainer's section on lorebooks and pins covers why keyword-triggered entries beat a giant always-on prompt, if you're new to this.)

Break one: Ottoline's card told on her

Back to Thursday, September 24. Here's why Ottoline confessed. Her Story Card triggered on her name and on the word “husband.” The card contained the secret. So the moment my inspector said “husband,” the whole truth landed in the model's context, labeled as a fact about her, and the model did what models do with facts in context: it used them.

My little instruction (“she lies about this until...”) was one line among many. The secret itself was vivid and specific. Vivid won.

The fix took ten minutes the next night. I split her into two cards. The public one (widow, runs the tide office, says her husband drowned in a spring storm) kept the ordinary triggers. The secret went into a second card that only triggers on “second ledger,” a phrase that couldn't come up until the inspector found it. After that she lied beautifully for the rest of the run, and when the ledger finally turned up on the second-to-last night, the confession arrived right on cue.

If you remember one thing from this review: a Story Card isn't a vault. Anything inside it is public the instant its trigger fires.

Break two: the ledger came back from the fire

The second failure was sneakier. On Saturday, September 26, the first ledger burned in the tide-office fire. Big scene. I'd written the Story Summary by hand up to that point and the fire was in it.

On Thursday, October 1, five days and maybe 130 actions later, Ottoline slid the first ledger across her desk and asked me to check the March entries. I opened the Story Summary. Auto Summarization had rewritten it. Latitude's own guide warns that with the feature on, the summary “will update itself over time and possibly erase or change what you wrote,” and that's exactly what it did: the new version mentioned a fire but not what burned. Everything downstream followed from that.

I moved “The first ledger was destroyed in the tide-office fire. Only the second ledger exists” into Plot Essentials, which I write and the summarizer doesn't, and it never slipped again. This is the same failure I hit in the summer campaign, where the rolling summary decided a flood was old news. Two stories, same weakness. I'm now fairly confident it's structural, not bad luck.

The Memory Bank did better than I expected. It recalled the night watchman's bad knee unprompted on day nine, a detail that was in no card and long gone from context. Latitude's What's New page says premium members got an “Enhanced Memory Bank” in August. I was on Journey for those last six days, and I can't tell you whether the knee came from the enhanced version or plain luck.

Twelve days, one harbor town, a 4.1

So, is it still the best free text adventure? Yes. Nothing else free gives you a narrator that understands the difference between doing and saying, lets bad choices hurt, and doesn't show you a single ad. It loses points for one thing, and it's the thing that matters most in a long story: the summary underneath will eventually rewrite your history, and you'll only notice when a burned ledger slides back across a desk. Keep the facts that matter in Plot Essentials and cards you wrote yourself, and it's the most fun I've had with a keyboard this fall.

It's for anyone who grew up on Zork or choose-your-own-adventure books, writers who want a sparring partner more than a ghostwriter, and people who bounced off companion apps because every conversation turned into a relationship. It's a poorer fit if you want one character who knows you over months (that's a companion app's job), or if keeping cards tidy sounds like homework, in which case the beginner roleplay apps will be more fun.

One more opinion, and plenty of AI Dungeon regulars will disagree: play your first week on the free tier even if you can afford Legend. The small free windows force you to learn which facts belong in a card and which in Plot Essentials, and that skill carries over to every tier and every other app.

Scenarios, the content dial, and one privacy toggle

You don't have to build a Saltmere. Most people start with a scenario: a premade opening, often with Story Cards and instructions already filled in, published by other players. The library is enormous and wildly uneven, the way any user-made library is. Two recent changes help. Since an August patch, scenario pages show a gallery of the scenario's Story Cards with search and spoiler covers, so you can judge how much work went into one before you play it. And as of a September 28 patch you get a quick option to rate a scenario after playing.

My test for a community scenario: open the card gallery. Fewer than three cards usually means a clever opening paragraph and nothing behind it.

Multiplayer is still there as well. Per the multiplayer FAQ, the host starts a game from the Play menu and shares an eight-digit join code, and there's no hard cap on players. I kept Saltmere solo this time.

Safe, Moderate, Mature

Content is controlled by a three-step setting. Safe blocks sexual, hateful, violent, or triggering content. Moderate reduces it. Mature doesn't limit it and asks you to confirm you're 18 or older. I ran Saltmere on Moderate, and the tide-office fire (a man with burns, a lot of smoke) came through with no friction at all. The App Store listing rates the app 13+, so a parent should know the 18+ step is a checkbox, not an ID check.

A bit of history explains why people still ask about safety. AI Dungeon launched in May 2019 as Nick Walton's GPT-2 project and moved to GPT-3 in 2020. In April 2021, Latitude added an automated filter that flagged innocent text and let moderators read flagged private stories, and the community revolted (the Wikipedia summary is fair). Today's player-set dial is a different system, but trust takes longer to rebuild than features do.

Check “Improve the AI” yourself

Latitude's privacy help page says the “Improve the AI” setting, which lets your story text be used to improve models, is off by default. Open your privacy settings after signing up and look.

AI Dungeon pricing in October 2026

Prices, memories, and credits come straight from the memberships page, checked October 4. The context column comes from the per-model figures on the model page, because the memberships page's single “up to” number hides how much it varies. Latitude also offers discounts for 6- and 12-month commitments, and says Journey and Legend include a one-week trial with purchase.

Above Ultimate sit three “Shadow” tiers (Banshee, Reaper, Apocalypse), each doubling the context of the one below on models that allow it. Latitude doesn't publish their prices on the help pages I could reach, and its own FAQ admits they “can be uncomfortable.” I'm not testing those on this blog's budget.

For a single story played a few evenings a week, Journey is the sensible stop, mainly because of what it does for the Basic models. Harbinger at 8K was worth the $14.99 on its own. Legend mostly matters if you want the big Premium models at more than 4K, or you're running a months-long campaign.

Next to NovelAI, SillyTavern, and the rest

A full rundown of AI Dungeon alternatives is coming next month, so this is the five-line version.

The pattern: everything that beats AI Dungeon on memory asks for more setup or more money, and everything that's free and easy is a chat app with a fantasy skin. My Character.AI vs SillyTavern comparison and the SillyTavern setup guide cover the build-it-yourself route; the DreamGen review and Perchance guide cover two of the others. If you came here from Character.AI specifically, the Character.AI alternatives list is the better starting point, since most of what's on it is chat-first.

Before you make an account

Ottoline is still in Saltmere, by the way. The second ledger is in my inspector's coat. Journey renews October 28, and I'll renew once to finish a 30-day story before deciding. That story starts tomorrow, from scratch, and I already know which of my Story Cards I'm splitting in half on night one. You can try it yourself at aidungeon.com, or grab the iOS or Android app.