DreamGen is a text-first AI role-play and story platform where a structured scenario keeps characters and world rules consistent, with paid tiers priced by context window rather than by output volume.
Used this tool? Rate it
Used this tool? Rate it
DreamGen is a browser-based platform for AI role-play and story generation in which the user defines the world and steers the plot while the model writes. It is text-first: there is no avatar rendering, no voice and no video, and the product is built around a structured scenario rather than an open chat box.
Role-play uses an interaction-based system that works more like a conversation, in contrast to story writing, which uses text blocks. The two directions share the same scenario definitions: a role-play scenario is switched into Story Mode at the top of the setup screen, and a story scenario is switched into Role-Play mode at the top of the screen, so the choice is about how you want to sit in the story rather than about which content you can build.
As an AI story generator it therefore competes on structure rather than on raw fluency. First-party models are hosted and fine-tuned in-house and split into Base and Premium tiers, sitting alongside third-party models, and the older Lucid Base and Lucid Max (Chonker) pair is now marked legacy and slated for deprecation. Running the models in-house is why the platform sets its own content boundary instead of inheriting one.
The structure that distinguishes this product from a chat interface is the scenario: a persistent definition the model consults on every turn.
A scenario carries eight defined elements — plot, setting, style, characters, locations, objects, opening and examples — and most of them are optional. That matters for the learning curve: a usable scenario can be one paragraph of premise, and the remaining fields exist for people who want tighter control over how the model behaves later.
The writing surface separates text blocks holding the story itself from instruction blocks that direct what the model does next, and everything auto-saves as you write. Instructions are inserted between passages rather than typed into a chat line, which is what lets you set up a scene change or a pacing target without that request becoming part of the prose.
Any passage can be edited, rewound or undone, including text the model produced, because the transcript is treated as your document rather than a chat log you merely received. Each session also gets its own copy of the scenario, editable mid-story, so a character who transforms halfway through can have their description updated without touching the original scenario for other sessions.
The context window has to hold the whole scenario definition plus the role-play generated so far, and when it will not fit, the oldest parts are truncated automatically. That ordering explains a behaviour people otherwise find puzzling: world rules stay reliable while mid-story details drift, because the definition keeps its place while the session's early stretches are the first thing dropped.
Individual interactions can be marked sticky, which gives them high priority in the AI's memory so they are more likely to remain in context as the story continues. That is the intended lever for pinning a plot revelation or a relationship change that the story needs to remember much later, though the documentation warns against overusing it.
Getting a first session running takes five steps.
Response length is steered per turn through an instruction that names an explicit word target — the documentation's own examples ask for a next message of at least 200 words or at least 100. Because it is an instruction block rather than a standing rule, it changes one beat of the story without changing its voice.
Instructions can control what happens next, how long a passage runs and how a character reacts — the documentation gives examples such as directing a flashback, specifying a two-hundred-word scene, or having one character reach for an object and open it.
When a response goes wrong, the documented move is to retry it — optionally with a different model — because both versions are preserved as branches you can navigate between rather than one overwriting the other. Switching models mid-session is treated as a normal adjustment rather than a restart.
These are configuration problems with documented answers rather than model failures you have to wait out. Knowing that changes how a bad session is handled: the overlong reply, the scenario that no longer fits the window and the response you simply dislike each have a named control attached rather than requiring a restart.
If the scenario definition itself stops fitting, the documented remedies are to shorten it or move to a model with a larger window, and the docs publish budgets to aim at: under 2,000 tokens on a 4,000-token window, under 5,000 on 8,000, and under 8,000 from 15,000 upward.
Which tier fits depends less on how much you write than on how long a single continuous thread needs to stay coherent.
Starter costs $7.83 a month for a 5,000-token window and 375 monthly credits, Advanced $19.35 for 15,000 tokens and 2,250 credits, and Pro $48.30 for 30,000 tokens and the same 2,250 credits. The credit allowance stops rising after Advanced, so the entire difference between the middle and top tiers is memory and discounts, not output volume.
That shape tells you who each tier is for. Someone running short scenes with a small cast rarely notices a 5,000-token ceiling. Someone maintaining a multi-chapter arc with a persistent cast hits it constantly, and pays for the window rather than for more words.
Community reports on the product's own forum describe memory holding up even on the unpaid tier and multi-character scenarios being handled better than on the other platforms those users had tried. That is directional sentiment from a small number of self-selected posters rather than measured behaviour, but it points at the same axis the pricing does.
The product runs in the browser and is reached at dreamgen.com along with associated apps referenced in the terms.
The primary generative models run on the vendor's own infrastructure, which is what makes the content policy self-determined rather than inherited from an upstream provider. Under conditions such as high demand or service disruption, or where a particular feature requires it, requests or related data may be routed to vetted third-party providers for text, audio, image or video generation, moderation, classification or embeddings.
No promise is made that the service or any specific model stays available, and models offered on an experimental basis are flagged as especially likely to degrade, change or disappear without notice. Anyone building a long project around one specific model should read that as a real planning constraint rather than boilerplate.
The model catalogue itself is split along the same line: first-party models hosted and fine-tuned in-house, grouped into Base and Premium tiers, with third-party models available beside them. Which of those you get without spending credits is what the subscription actually buys, since a plan may unlock the Base tier or every text model.
The paid ladder is priced by memory rather than by output volume, which is unusual enough to change how you read it.
| Plan | Per month | Context window | Monthly credits | Extra daily credits |
|---|---|---|---|---|
| Free | $0 | 4,000 tokens | limited | limited |
| Starter | $7.83 | 5,000 tokens | 375 | 45 |
| Advanced | $19.35 | 15,000 tokens | 2,250 | 75 |
| Pro | $48.30 | 30,000 tokens | 2,250 | 75 |
A genuine free tier sits below all of them, opening access to every current model on a limited monthly and daily credit allowance and a basic 4,000-token context window, enough to judge the writing before committing. Credit consumption for text depends on the model, on how long the session already is in tokens, on the length of the response and on the plan, and monthly usage resets on a day derived from your subscription start date rather than on a fixed calendar boundary.
The daily allowance scales too, listing 45 extra credits a day on Starter and 75 on both Advanced and Pro, but it is a safety net rather than a bonus: it only activates once the monthly limit has been used up, and it does not carry over from one day to the next. That makes the monthly figure the real budget and the daily one a floor under it.
Only Pro carries a 20 percent discount on credit packs and up to 60 percent off selected third-party models; both cheaper paid tiers list zero on each. That structure is what separates this from platforms priced by message count, and it is the axis worth comparing on rather than the monthly figure. A tool billing per message rewards short exchanges; this one rewards a long thread you keep coherent, and the two only look comparable until you count tokens.
Stories and role-plays are allowed to grow longer than the context window, and when they do the model simply cannot see all of the material at once while generating. This is the single most consequential boundary in the product and it is documented rather than hidden: no plan removes it, larger windows only push it later.
Access is restricted to people who are at least 18 and of the age of majority where they live, and anyone who does not meet that bar is prohibited from using the service. The content boundary is drawn at anything illegal in the United States, the EFTA countries or your own jurisdiction, plus any sexual content involving minors, fictional minors included.
Review and removal of content are described as a right the vendor holds rather than an obligation it accepts, exercised at its sole discretion through both automated and manual systems. Automated moderation decisions stay unreviewed by any person unless you appeal, unless the content was published or submitted for publication, or unless a legal obligation forces review.
Purchases are final and non-refundable except where consumer protection law demands otherwise or the vendor chooses to refund at its own discretion. Disputes fall under Swiss law and binding arbitration seated in Zurich, with a class-action waiver attached.
An independent review blog puts the free tier's context window at 4,000 tokens with a fixed monthly credit allowance and no card required, matching the figure the vendor's own pricing page states. The free window is therefore roughly a fifth of Starter's and an eighth of Pro's, which is what makes it a trial rather than a working tier.
The service claims no ownership or copyright over what you generate, but it also makes no representation that AI output is copyrightable under applicable law. Those are two different statements and both matter: nobody is taking your story, and nobody is guaranteeing you can stop someone else from using it.
You grant an unrestricted, worldwide, transferable, sublicensable and royalty-free licence over your content for operating, providing and improving the platform. Training consent is off by default and the vendor states plainly that scenario definitions and session history are not used to train or improve models unless the user opts in. Withdrawing that consent stops future use but does not pull already-collected material back out of existing training datasets or model weights.
Termination clears data from active systems within a reasonable period, but some may be kept for legal, accounting, safety, backup or model improvement reasons, which is why exporting first is the vendor's own recommendation.
Yes. A free tier opens every current model on a limited monthly and daily credit allowance with a basic 4,000-token context window, enough to try the features before paying. Monthly usage resets on a schedule anchored to your account creation or last subscription start, not to the first of the month.
Mostly the context window: 5,000 tokens on Starter, 15,000 on Advanced, 30,000 on Pro. Monthly credits stop rising after Advanced, so the jump from Advanced to Pro buys memory and discounts rather than more writing.
Because the window has to fit the scenario definition and the whole role-play so far, and when it will not, the oldest parts are truncated automatically. Older details fall outside that window unless you mark the interaction sticky or fold the fact back into the scenario.
Retry it, optionally with a different model. Both versions are kept as branches you can move between with the arrow buttons, so nothing is lost and you can compare continuations instead of committing to the first one.
Not unless you opt in. Consent lives in account settings and is off by default. If you do opt in and later withdraw, future use stops but material already used is not retroactively removed from datasets or weights.
The service claims no ownership or copyright over your content, but it explicitly makes no representation that AI output is copyrightable under applicable law, and you grant the platform a broad licence to use your content in operating and improving the service.
Shorten it or switch to a model with a larger context window. The documentation publishes target budgets: stay under 2,000 tokens on a 4,000-token window, under 5,000 on 8,000, and under 8,000 once the window reaches 15,000 or more.
Purchases are final and non-refundable, with two exceptions: where applicable consumer protection law requires a refund, and where the vendor chooses to grant one at its own discretion. The vendor's pricing page says the opposite — that prorated refunds are offered, requested by email, with the amount set by days remaining, credits used and prior refund history, and most processed within two business days. The two pages have not been reconciled, so treat the outcome as unsettled and ask before paying.
