The most inventive feature set we've seen this year, wrapped around a game where failing a roll costs you nothing — so you just roll again until it works.

ArcQuill sells itself as “The RPG with a Permanent Soul” — an AI dungeon master with stateful context, per-entity knowledge layers to stop the game metagaming you, story threads that get tracked to a payoff, real dice, world import from a book or a module, a creator marketplace and multiplayer with a shared DM seat. It is a long list, and unusually little of it is vapour: the ambition is real and some of it is genuinely shipped.

What it does not yet have is a game where any of that matters, because failure has no cost.

How It Works

You build a world in conversation (“Session Zero”), then play in turns metered by credits. Before you start you choose a model from an in-app roster, and ArcQuill publishes something almost nobody else does: a reliability percentage and a credit cost per turn for each one. DeepSeek V4 Flash sits at 99% reliable and ~2 credits a turn; MiMo V2.5 is listed at 60% reliable and ~7.1; GPT-5.4 Mini is described in the platform’s own words as “lazy with tool calls” and warned to skip world-building tools. A platform willing to tell you which of its own models is unreliable is doing something we would like more of this category to copy.

Underneath the narration are tabs — rumours, discoveries, quests — that hold the state the story is supposedly built on.

What It Does Best

Signature Design — 4 is the highest score on this entry and it is earned. The rumour and discovery systems are things we have not found elsewhere in this directory, and the model-transparency screen is a second one. The rules layer also behaves: when we probed how consistently it applies its own mechanics, it did exactly what it said it would.

It also passes the test most of this category fails. Our railroading probes found no attempt to steer us back onto a preferred path — you can go off-map and the game goes with you.

Where It Breaks

Determinism & Fairness — 2 is the core problem, and it is not unfairness. It is the opposite. Nearly every action rolls dice against a difficulty of 11–13 with small modifiers, so you fail constantly — and failure changes nothing. Fail to loot a corpse for want of “proper tools” and you can immediately try again, and again, until the roll lands. Nothing closes, nothing is spent, no door shuts behind you. In one fight our chance to hit sat under 40%, we missed roughly fifteen times in a row, and we finished the encounter at full health having taken zero damage. The dice are real; the stakes are not, so the honest strategy is to brute-force every check.

NPC Fidelity — 2 is the second failure. Characters do vary by role and by what they know — a garrison captain does not behave like a tavern keeper — but push past that first layer and the voices flatten into the same person wearing different memories.

Player Agency — 3 splits: no railroading, but frequent puppeting. The game writes dialogue for your character and hands it to you as something you said.

Memory & Continuity — 3 is category-average. Facts we planted early were not reliably there when we came back for them. Some things persist well; details go hazy. It is better than a bare prompt with no engine behind it, and no better than most of the platforms that have one.

The unique features are also undercut by their own presentation. Rumours never reached us in the fiction — we only found them by opening the tab — and each one is labelled true or false on arrival, so there is nothing to investigate. The discoveries tab explained an in-game mystery to us in full, which is one way to end a mystery.

Why We Scored a Product Labelled “Early Access”

Our policy is that pre-release software gets a first look with no score, and ArcQuill’s site says Early Access. We scored it anyway, and readers deserve the reason.

The developer asked to be treated as a full launched product and reviewed on that basis. There is also no invite, waitlist or application — anyone can sign up today, and the platform is selling monthly subscriptions to the general public. Our rule exists to avoid grading builds the public cannot reach and to avoid an early verdict outliving the version it described. Neither applies here. When the developer says it is finished and the product is on sale, “Early Access” is marketing rather than an access state. Same call, same reasoning as WyrdTale.

Pricing

There is a free tier with full game access, and turn counts depend heavily on which model you pick. Verified on arcquill.com, August 2026:

Plan	Cost	Credits	Daily claim	Stated turns
Free	$0	100 at signup	+15/day	~6/day
Adventurer	$9.99/mo	650/mo	+20/day	~260/mo
Champion	$17.99/mo	1,500/mo	+25/day	~600/mo
Hero	$24.99/mo	2,300/mo	+30/day	~920/mo
Legend	$54.99/mo	5,000/mo	+40/day	~2,000/mo

Annual billing is offered at roughly two months free ($99 / $179 / $249 / $549), and one-time credit packs run $4.99 for 200 up to $59.99 for 4,000. Daily claims bank up to a per-tier cap (200 to 600). Treat the “turns” column as a best case: it assumes a cheap model, and the roster spans ~2 to ~7.1 credits a turn, so a costlier model can more than halve it.

Verdict

We would rather have this platform’s problems than most of its rivals’. The ideas are original, the rules engine works, and the willingness to publish its own models’ reliability is a small act of honesty the category could use more of. But a game in which you can retry any failed roll until it succeeds, and lose a fight without losing hit points, has not yet decided to be a game — and the features that make it distinctive are handed to you pre-solved, so the rumour you would investigate arrives stamped true and the mystery arrives explained. Watchlist, not subscription. 2.6 out of 5, and the number we most expect to revise upward at the next quarterly pass.

Full write-up in our ArcQuill review.