Exclusive: Voyage's Open Beta Is Weeks Away — Latitude CEO Nick Walton on the 2027 Release Date, the Altar Problem, and the Axis Nobody Has Solved

Access disclosure: Latitude provided beta access to Voyage. No conditions were attached to it or to this interview — this coverage is independent and unpaid.

We sat down with Nick Walton, CEO of Latitude — the company that made AI Dungeon in 2019 and has spent the last five years building its structured successor, Voyage. We came in with questions about architecture, not marketing: where the deterministic engine stops and the language model starts, what six prototypes taught them, and which of the unsolved problems on our benchmark he thinks is genuinely the hardest.

A note before we start. Arcanum is an independent site. We have no affiliation with, endorsement from, or sponsorship by Latitude — we name AI Dungeon and Voyage here to report on them.

Open Beta Is Weeks Away. Full Release Early 2027.

The headline first, because it’s the thing players have been waiting on.

Voyage has been invite-gated since Latitude unveiled it in April 2026, with an open beta vaguely scheduled for “later in 2026.” We asked what bar it has to clear before Walton would call it ready.

“Open Beta will be, probably, in weeks.”

That’s a significant tightening from anything Latitude has said publicly, and it’s the practical takeaway for anyone currently sitting on a waitlist: the invite code stops being the barrier shortly. Full release is a different matter, and it’s gated on marketing, communication, and quality rather than on any single feature landing: “We are expecting the full release by early 2027.”

We’ve updated our Voyage review and Voyage vs AI Dungeon to reflect both dates. It also means Voyage stops being exempt from our benchmark soon — we don’t score pre-release software, and that clock is now running.

The rest of the interview is about what you’ll actually find when the gate opens.

A Lenient Narrator, and Nothing to Be Afraid Of

We opened by setting the product aside and asking which problem annoyed him as a player.

His answer was about danger. What he wanted was a persistent world with real challenge, and AI Dungeon couldn’t give him one. He described its narrator as “a lenient narrator” — a game master that says yes to everything, including things the player has simply asserted into existence. Players could hallucinate items, allies, and escapes, and the story would go along with it.

The consequence isn’t just an occasional broken scene. It’s structural: with no possibility of failure, there is no threat, and with no threat, objectives stop meaning anything. You cannot build tension on top of a narrator that will always let you win.

This is the exact frustration we described in our Voyage review from the outside — the slow realization that nothing you do matters. It’s worth noting that the person who built AI Dungeon arrived at the same diagnosis from the inside, and it’s the reason Voyage exists at all.

The Altar Problem

Six prototypes over five years. We asked what the first one got wrong and what specific insight made the sixth one work.

Walton was blunt that most of them simply didn’t work, and a large part of that was the models available at the time. “The GPT-2 era wasn’t good enough for Voyage, and while GPT-3 was better, the turning point was GPT-5.” That’s an unusually direct admission that a product’s viability was gated on external model capability rather than internal cleverness — and it explains the five-year gap better than any roadmap would.

But the design insight is the part worth stealing, and he has a name for it: the Altar Problem.

Here’s the shape of it. You open a room in a dungeon. It contains an altar and a door. You walk through the door — and inside is a room with an altar and a door. The generator is producing locally plausible content with no global logic. Every individual room is fine. The world they add up to is nonsense.

The Altar Problem is what happens when you generate a world piece by piece, on demand, with the model only ever seeing the immediate frame. Nothing in that loop knows what the dungeon is, so nothing can stop it from recursively re-describing the same room forever.

Latitude’s answer was to invert the order of generation. Rather than building the world incrementally as the player walks into it, Voyage generates holistically — the structure exists first, and detail is filled in afterward. On top of that sits something Walton described in terms borrowed from graphics: rendered levels of detail. A dwarf seen across a valley is just a dwarf. Walk up to him and the narrative generation resolves him into a person with a name, a personality, objectives, and gear.

That’s a genuinely good idea, and it’s the most transferable thing in this interview. It reframes the AI RPG memory problem as a resolution problem instead of a storage problem. You don’t need to remember every detail of every entity in the world, because most of them shouldn’t have details yet. Detail is something you pay for on approach.

It also quietly explains why the holistic-first approach is a prerequisite rather than a nicety: you can only render a dwarf at low detail if something already knows there is a dwarf.

Where the Engine Stops and the Narrator Starts

We asked exactly where the line sits between the deterministic systems and the AI narration — what does the engine decide, and what does the model decide?

Voyage resolves a turn in phases. A pre-story phase determines what NPCs are going to do, whether the player succeeds at what they attempted, what items they actually have, how successful the attempt was, and how much damage lands. Combat is simulated by the engine outright.

Only then does the model narrate — and it narrates under constraints handed to it by that simulation.

Walton’s framing was that the model narrates while the engine underneath does the brunt of the work. That’s the opposite of how most of the category is built, and it’s the reason Voyage can have permadeath that means something. If the outcome is decided before a single word is generated, the words cannot talk their way out of it. The narrator is a renderer, not a judge.

For anyone building in this space, that phase order is the whole design. Resolve, then narrate. Most platforms narrate and then try to reconcile.

When They Disagree, There Is No Rule

We pushed on the obvious failure case. The engine says the sword was lost three scenes ago; the model narrates the character drawing it. Who wins, and how do you enforce that?

This is where we got the least satisfying answer of the interview, and we’re reporting it that way because the honesty is the point. Walton said it depends on where the boundary sits. State can hold something the model doesn’t pick up — and the reverse happens too, where the narration is right and the recorded state is the thing that’s wrong. He didn’t offer a single arbitration rule, because he doesn’t think there is one.

We’d have liked a harder answer here. But “there is no universal winner” is at least a real position, and it’s more credible than claiming the engine is always authoritative — a claim that would be trivially falsifiable the first time a player saw a technically-correct-but-nonsensical scene.

It does leave the most interesting question open. Determinism is only as good as the state it’s deterministic about, and nothing Walton described adjudicates a disagreement automatically. That’s a seam worth watching once open beta puts real players against it at volume.

Structure vs. Magic: The Narrator Agent

There’s a real fear among developers that bolting a state engine onto an AI RPG costs you the improvisational surprise that made AI Dungeon feel alive — that you trade magic for bookkeeping. We asked whether Latitude hit that tension.

He acknowledged it directly. The tension is real. Voyage’s answer is a narrator agent — a window of support that can change things on the fly, letting a player steer the story toward what they want it to be.

Having used it, we’d say that’s accurate rather than a talking point: when the fiction needs to move somewhere the simulation didn’t plan for, there’s a mechanism to move it. It’s a deliberate escape hatch in an otherwise deterministic system, which is a more sophisticated position than either extreme. Total fiat produces AI Dungeon’s lenient narrator. Total rigidity produces a spreadsheet.

Whether the hatch is sized right is the open question, and it’s the one we’ll be testing when we eventually rate the platform.

Player Agency: The Axis AI Dungeon Still Owns Outright

Our benchmark scores every rated platform on seven axes. Three of them have no top score anywhere in the field — memory and continuity, player agency, and longevity. Nobody has hit 5 on any of them.

On player agency specifically, AI Dungeon is alone at the field ceiling. It’s the only platform on the board that gets there, against a field average of 2.25. So we asked what Latitude did differently.

The answer was models, not architecture: custom models trained specifically to respect player agency and to produce fewer clichés and tropes. Models trained, as he put it, to solve the problems commercial models have consistently struggled with.

That’s consistent with what we’ve observed. AI Dungeon’s Player Agency — 4 is not a product of clever orchestration; the platform barely has any. It comes from a narrator that doesn’t seize control of your character and doesn’t railroad you back onto a path — and Latitude got there by fine-tuning for that behavior rather than engineering around it. It’s also a reminder that agency and structure are separable problems. AI Dungeon has the field’s best agency score and one of its weakest mechanical depth scores, and the same company is now trying to hold the first while fixing the second.

Longevity Is the Hardest — And He Didn’t Hesitate

We asked which of those three unclaimed axes is genuinely hardest, and whether it’s a model problem or an architecture problem.

“Longevity is the hardest.”

He answered immediately and moved on; the conversation went elsewhere before he elaborated, so we can’t tell you which side of the model/architecture split he’d land on. But the speed of the answer is itself informative, and it matches our data. Longevity has the second-lowest field average on the board at 2.31, three platforms score at or below 1, and nothing has ever cleared 4.

It’s the axis that asks whether a platform survives past turn 50 and the novelty window — and it’s the one where the entire category, Latitude included, is still guessing.

Breaking Up With OpenAI, Then Taking Google’s Money

AI Dungeon started on OpenAI and now runs on a mix of open models — Llama, DeepSeek, Gemma, GLM. We asked what going model-independent cost, and what it bought.

He called it the best move they made. When AI Dungeon launched there was nothing to choose from; GPT was the only option, and that dependence had a cost. Independence bought them control, privacy, moderation on their own terms, the ability to train their own models, and lower costs.

The obvious follow-up: Voyage leans on Gemini Flash and Gemma, with Google’s AI Futures Fund behind it. Having just spent years unwinding a dependency on one provider, how do you avoid rebuilding the same exposure?

His answer was that the deal is structured to avoid exactly that — they can switch models, and there’s no obligation to use Google’s.

Worth taking at face value, with the caveat that contractual freedom and practical freedom aren’t the same thing. A platform tuned around one provider’s latency and pricing tends to stay there regardless of what the contract permits. But the lesson does appear to have been learned at the level where it counts, which is the paperwork.

The 2021 Filter, and What Moderation Should Actually Be

In 2021, AI Dungeon’s content filter — and the disclosure that human moderators could read private stories — did real damage with its community. We asked what that changed about how they build, and what’s structurally different in Voyage.

He identified two root causes. The first was the dependence on OpenAI, which no longer exists; that autonomy is now real. The second was policy, and he was direct about it: “Second thing is what we allow, what we don’t allow, and what moderation should look like. OpenAI tried to apply social media moderation to something that should be private.”

That last sentence is the most quotable thing in the interview, and it’s a diagnosis worth sitting with. The 2021 blowup is usually remembered as a filter that was too aggressive. Walton’s framing is that the filter was the wrong category of tool — moderation designed for content broadcast to an audience, applied to a single-player story nobody else was going to read. The mismatch, not the strictness, was the failure.

Whether Voyage’s practice matches that principle is something only time and its open beta can demonstrate. But it’s a coherent stance, and it’s a materially different one from “we’ll be less strict this time.”

Two Pricing Pages, and the Tiers They Hide on Purpose

We noticed something odd: Voyage has two different pricing pages. One is reachable from the main site; another only appears once you’re inside your account. We asked what that’s about.

The answer was more candid than we expected, and the structure underneath it is worth understanding, because it’s unusual.

One subscription covers both products. The publicly visible ladder is a single Latitude subscription spanning AI Dungeon and Voyage — Journey at $14.99/month, Legend at $29.99, Mythic at $49.99, and Ultimate at $99.99. You are not buying Voyage separately from AI Dungeon. You are buying access to Latitude.

Above that sits a second, unadvertised set of plans, and these cover both products too — the same one-subscription logic, simply not shown on the pricing page. Latitude calls them Shadow plans, and they are not modest:

Shadow plan	Monthly	Listed models & context	Memories
Banshee	$246.66	64k DeepSeek / 32k Raven	1,200
Reaper	$496.66	128k DeepSeek / 64k Raven	1,600
Apocalypse	$996.66	128k Raven	2,000

Each also carries a monthly credit grant for image generation, a Shadow-exclusive channel on Latitude’s Discord, and “Shadow Clan” — a family-plan mechanic where each tier grants slots you spend gifting premium access to other players. The context figures apply on the AI Dungeon side, doubling at each step; Voyage doesn’t tier context at all, and Latitude has said it doesn’t plan to.

Walton’s reasoning for their existence was simple. People are happy to pay more for a better experience, and these plans exist because players asked for them.

The reason they’re hidden isn’t embarrassment. It’s that leading with a very high number teaches prospective players that they need to spend that much for the product to work properly — which isn’t true, and which would scare off people the standard tiers would have served fine.

Our own read: a $996.66/month tier is the highest consumer price point we’re aware of anywhere in this category, and the fact that it’s demand-driven rather than promoted is the interesting part. It says there’s a segment of players for whom context length is the binding constraint on their campaign, and who will pay nearly a thousand dollars a month to lift it. That is a much sharper signal about the memory problem than any benchmark score — some players are effectively buying their way around it.

We asked separately whether pricing is driven by inference cost or by positioning. Mostly inference cost: Voyage needs to be cheaper at full launch than it is now, and Latitude is betting on model prices continuing to fall to get there.

That’s a real bet with a real failure mode, and it’s the one commercial risk in this interview that isn’t hand-waved. If inference costs don’t fall on schedule, the launch pricing either rises or the margins don’t work.

What the Whole Category Still Gets Wrong

Our closing line of questioning: what does this category consistently get wrong that you wish someone would just fix?

His first answer was cost optimization across the industry — and he thinks it will get fixed, which is consistent with the pricing bet above.

The more interesting part was what he expects will not get fixed. Narrative quality and player agency aren’t going to reach Critical Role levels any time soon. And then the sharpest observation in the interview: the category is good at sandboxes and bad at direction. “Having games where you feel there’s a clear goal and challenges” still doesn’t happen. Multi-chapter stories that build toward a final goal are, in his words, almost unseen.

We’d go further — this maps almost exactly onto the longevity problem he named as hardest. A sandbox has no reason to end, which sounds like a strength until you notice it also means it has no reason to continue. Platforms don’t survive past turn 50 because there’s nothing at turn 50 that the player is trying to reach. Directed structure and longevity may not be two problems at all.

One correction on our own number, since the term is ours. “Turn 50” comes from fifty-turn amnesia, a label we coined back when campaigns reliably came apart right about there — and it stuck, including in the wording of our own longevity axis. It’s dated now. The engineering has genuinely moved: the better platforms hold coherence far longer, and the figure we watch today is approaching turn 500.

But moving the wall is not the same as removing it, and Walton’s point survives the correction intact — arguably it gets sharper. A longer sandbox is still a sandbox. Whether the fiction stops making sense at turn 50 or turn 500, a campaign with nothing to reach doesn’t end when the context window fills; it ends earlier than that, when the player stops caring. That’s the failure mode a bigger context window cannot fix, and it’s why “longevity is the hardest” is a design answer rather than a memory one.

What Failure Looks Like

We closed by asking what would make him look back in two years and call Voyage a failure.

“If it doesn’t give people great stories. People excuse visuals, people excuse mechanics, people don’t excuse bad stories. For RPGs, story comes first, mechanics come second.”

Coming at the end of an hour largely spent discussing deterministic combat simulation, phase ordering, and state arbitration, that’s either a contradiction or the entire thesis. We think it’s the thesis. Every piece of machinery Walton described exists to make the narration trustworthy — the engine isn’t there to replace the story, it’s there to stop the story from lying. The dice don’t matter. What the dice make possible is a scene where the outcome wasn’t guaranteed, which is the only kind of scene worth reading.

What We Take From This

Three things stand out.

The Altar Problem is the most useful concept anyone in this category has named this year, and the holistic-generation-plus-levels-of-detail answer is a legitimately different approach to a problem most platforms are still attacking with bigger context windows.

The phase order — resolve, then narrate — is the architectural commitment that everything else in Voyage rests on, and it’s what separates real consequences from narrated ones.

And the unanswered question is arbitration. Walton was honest that there’s no universal rule for engine-versus-narrator disagreements, which is the seam most likely to show under load. That’s what we’ll be looking at hardest.

Voyage still carries no rating from us, per our methodology — we don’t score pre-release software, and an invite-gated beta can’t be fairly measured. That changes when the open beta lands, which is now a matter of weeks. When it does, Voyage enters the benchmark like everything else, and we’ll find out whether an engine that decides outcomes before the narrator speaks can finally put a 5 on one of those three empty axes.

In the meantime: AI Dungeon is open to everyone today and still holds the field’s best player-agency score — our full review covers where it holds up and where it doesn’t. If you want the structured, consequence-heavy side of this argument, Voyage vs AI Dungeon breaks down what actually separates them. And if you’d rather play something tonight without a waitlist, our directory of AI RPG platforms covers everything currently open to the public.