DreamGen AI Review: The AI Co-Writer That Got My Novel Out of the Drawer

If you came for a girlfriend who texts good morning, you are in the wrong building. If you came to write, pull up a chair. That is the whole verdict in two sentences, and the rest of this review is the evidence.

My own run ended in a way I didn't expect from a chatbot. I paused the subscription at 10 weeks, when wedding season started booking up my weekends. The noir-ish novel I've been threatening to write since college is out of the drawer and on my desk, and I credit this app for that, not for any single sentence it wrote.

Somewhere in those weeks I learned that DreamGen trains its own 'Opus' storytelling models and puts some of them on Hugging Face. The tinkerer in me respected a company that shows its homework. As of September 2026 it runs its own fine-tuned, self-hosted first-party models.

The older 'Opus' open-weights models are still published on Hugging Face.

A third-party review sums it up as no voice, no face, a learning curve and a free tier that runs dry fast. I agree with all four points. I'd still recommend it to anyone who writes, because none of those four things is what it is for.

Key takeaways

• It is a writing tool with role-play attached, not a companion, so judge it as a co-author and not as a partner.
• Steering is the best part: plain instructions change the tone of the prose, and the scenario tools keep a world consistent.
• The learning curve is real, and the free tier's small context is too tight to judge the product fairly.
• Credits are a second currency on top of the subscription, and pictures spend them faster than words do.
• Privacy is a strength: a Swiss company, opt-in training, and clear labels on outside models that may train on your text.

I wanted a narrator I could argue with, and Story Mode gave me one.

That choice says more about me than it does about the app. I spend my working days deciding what stays in a cut and what goes. Handing the pen to a character and reacting to it felt like being a guest in my own edit.

What came out of it

By week 8 the model and I had 11 pages of usable draft. Maybe 30% of that survived my editing, which is roughly my survival rate for raw footage too.

That number sounds like a complaint, and it isn't. On a shoot most of what the camera records never reaches the client, and nobody calls the camera a failure. The job is choosing.

What the app gave me was volume with a point of view. Every evening there was new material to cut, and cutting is the part of writing I know how to do. A blank page has always scared me more than a bad page, and DreamGen made sure I never had a blank one.

The ending is in the verdict above: wedding season won. There was no fight with the app and no billing drama in how it ended. The table under this section lists the plain facts, and the rest of the review explains how each one felt.

By the time the scenario looked finished, it was too late to write anything worth keeping.

Part of me enjoyed it, the same part that reads camera manuals for fun. The other part had come to write, and it sulked until bedtime.

It reminded me of the first day on new editing software, when every panel is open and none of them is yours yet. Nothing was broken. It was just built for someone who already knew where everything lived.

Looking back, that evening was an investment, not a waste. The scenario I built that night carried the story for weeks. I'd still warn anyone who wants a quick chat after dinner: DreamGen asks for homework before it gives anything back.

How sign-up looks today

The sign-up notes below come from a check on 23 September 2026, which used the web version on a free account. The site is live. DreamGen runs on the web only, with an English interface.

There are two addresses to know. dreamgen.com is the marketing site and blog. The app itself runs at v2.dreamgen.com, and the sign-in links redirect there.

Sign-up offers Google OAuth or e-mail and password. E-mail sign-up needs no username. A username can be set later in Profile, and the page warns that it "can't be changed later".

I'd leave it empty until you know what you want to be called, because this is one decision the app won't let you revise.

Google sign-in is the quicker door. E-mail is the better one if you'd rather keep a writing app apart from the account that holds your calendar and photos, and the verification step below is quick enough not to matter.

Verification and the tick boxes

E-mail verification is required. A code-and-link e-mail titled "DreamGen: Verify Your Email Address" comes from [email protected]. It arrived in under 1 minute. The code and the link are valid for 12 hours.

Clicking the link pre-fills the code, but the "Verify email" button still has to be pressed. It is a small thing, like a render that finishes but still wants you to click Save.

After verification a "Before You Continue" box asks for 2 tick boxes. One accepts the Terms of Service, and the other confirms "I am 18 or older and a legal adult in my jurisdiction". There is no date-of-birth entry.

I'd call that an honest gate rather than a strict one. It asks for a promise and records it, and the content-settings section covers what that promise unlocks.

The disclaimer is worth reading once. It tells you the platform sees itself as a fiction tool, and that what you steer it toward is yours to answer for, which is how a writer should think about it anyway.

That is fast for the paperwork. The real first night starts after it, in the scenario tools, and that part still takes an evening.

The sidebar is the real front door. It lists Public Scenarios (Role-Play/Story), Your Role-Plays, Your Stories, Your Scenarios, Collections, Image Generate, Building Blocks, Scenario Wizard, Import and Docs. That is a lot of doors for a first visit, and I think it tells you who the app was built for.

What is on the shelves

Public Scenarios is a grid of user-published scenarios with a search box and filters. Token-heavy scenarios carry a "Long" badge. I like that badge more than it deserves, because on DreamGen a long scenario eats more of your context and your credits, and the badge warns you before the meter starts.

If you didn't, it can feel like walking into a convention without a badge.

Romance and adult-leaning cards also appear, among them a roommate story and one titled "Your Girlfriend Amy". They are a minority next to the adventure settings, which again says the crowd here is mostly writing plots, not partners.

What is hard to find

What I'd want from the library is a shelf for original, grown-up fiction: a mystery, a slow literary drama, a lighthouse story. The search box helps if you know what to type. Browsing, though, is dominated by the franchise crowd.

My advice is to treat the library as a set of examples rather than a menu. Open a few scenarios to see how other people structure a plot, then build your own. That is where DreamGen is strongest anyway.

That last part is what sold me. In editing, continuity is the thing nobody praises and everybody notices when it breaks. The codex was my continuity supervisor, and for most of the story it did the job better than some humans I have worked with.

What was called the Scenario Codex in 2024 is the scenario editor in the current V2 app. It covers Plot, Setting, Style, History, entities (Character, Persona, Location, Object, Narrator), Openings, Examples, Images and Private Notes.

Three ways to start

Scenario creation offers three routes: New from scratch (Role Play or Story), the Scenario Wizard (chat-based brainstorming with example prompts) and Import (exports or third-party character cards). The "Scenario Wizard" is an AI assistant that helps build scenarios. Import takes exported sessions or third-party character cards.

A reviewer rates the scenario structure and the Scenario Wizard as the main strengths, with a learning curve for new users. I agree on both counts, and I'd start with the Wizard if the blank editor looks like a tax form.

Personas

A new session asks for a persona: a name and a description. The session also shows the creator's suggested persona template, which asks for an age field marked adults only.

I think the persona matters less in Story Mode, where you are the author rather than a character. In role-play it is the difference between a stranger wandering into the scene and a character the story can react to, so it deserves two honest sentences.

For the notes in this section, a blank private role-play scenario draft was opened, and nothing was published. The editor is still dense, patient and ready to reward whoever fills it in properly.

That is the feature I'd put on the box if I ran the company. Most chat apps let you write the next line. DreamGen let me give a note, the way I give notes to a colourist, and the next page came back graded the way I asked.

It changed how I worked. Instead of rewriting a flat paragraph, I asked for the mood I wanted and cut from what came back. It felt less like typing and more like directing a very fast second unit.

The purple patches

The prose sometimes gets purple. I got 'the sea sighed its ancient sigh' three times in one chapter, and I left margin comments like client revision notes out of habit.

In fairness, a sighing sea suits a lighthouse story once. By the third time it was a continuity error with a thesaurus, and my notes started to sound like the ones I send about a logo that keeps changing colour.

The default model is GLM 5.2. Each reply took about 15-20 s. That wait is fine for prose and slow for banter, like a satellite delay on a live interview.

Replies were long and literary: a separate "(narrative)" block plus a message from the NPC "Maren", the owner.

The bot quotes are the best evidence of the tone.

One reply named the shop cats: "Mr. Darcy is aloof but reliable. Sgt. Peppers is friendly but will knock things off shelves if he's feeling neglected."

Another described a neighbour: "The woman who runs the bakery two doors down? Writes the most devastating sonnets about bread." That is a character with a life outside the scene, which is exactly what a bookstore story needs.

The prose quality is good, with few clichés. I'd rate that line about sonnets above half the dialogue in the romantic comedies my job has made me sit through.

A session can hold narrative blocks and multiple NPCs. The homepage advertises the same thing: "Talk with multiple characters at once or explore fantasy realms."

The controls around the chat

Under the chat, "Continue As" offers You, NPC or Custom: "Continue As: You / NPC / Custom". "Suggest Next Steps" offers choose-your-own-adventure options. A CYOA "Choose the Next Step" mechanic turns those into clickable options.

The Continue As buttons are the underrated part. Letting the NPC take the next turn, or writing a custom one, keeps a scene moving when the ideas run out late at night.

The right panel has Scenario (Private Notes, Plot, Style, Setting, History, Personas, Entities, Examples, Images), Settings and Assistant tabs. Session settings cover the model, max input tokens, max output tokens, max output interactions, Thinking Effort, Take Turns, Narrative Enabled and Next Steps (CYOA) options.

There is also plenty of UI theming: avatar shape and size, font, text, italics, bold and quote colours, and background. Settings include Debug and "Set as Default" too. It is the cockpit of an app that assumes you will fly it every night.

The model menu

The first-party models in the picker are GLM 5.3 (experimental), GLM 5.2 ("Best first-party model"), GLM 4.7 ("Balanced choice"), Gemma 4 31B, Gemma 4 26B A4B and Qwen 3.5 27B. Those are not the 'Opus' models of 2024, which is the normal fate of anything in this field that is two years old.

A large menu of third-party models sits beside them: DeepSeek V4, GLM 5.x, Kimi K2.6, MiniMax M3, MiMo, Muse Spark and Glimmer, and Inkling. In full, the third-party options include DeepSeek V4 Pro ("One of the best models"), DeepSeek V4/V4.1 Flash, GLM 5/5.1/5.3/5.3 Flash, Kimi K2.6, MiniMax M3, MiMo V2.5 Pro, Muse Spark 1.2/1.3 (plus Contributor), Muse Glimmer, Inkling and Inkling Small.

I value that label more than the model names, because it tells you who else might read your chapter.

For a writer the practical advice is short. Start on the default, draft on a cheaper model, and switch back when a scene matters, the way you would render proxies for the rough cut and full quality for the export.

My view after all this: the conversations are good, the controls are better, and the best way to use them is as a director, not as a pen pal.

Where Ada lost the plot

Context drift arrives around message 20–30 even on paid plans. Ada forgot her brother had drowned and invited him to dinner. It was eerie for about a second, then just wrong.

For a second I read it as a ghost story beat, and I liked it. Then I remembered I hadn't written a ghost story, and it became what it was: a continuity error. In a film edit that is the actor wearing a different jacket after the cut, and in a novel it is worse, because the reader stops trusting the narrator.

My fix was the one every editor knows: keep the important facts where the model can't miss them. The scenario tools are built for exactly that, and they helped far more than hoping the model would remember.

What the plans promise

Desk research of the official pricing page listed context windows of 4,000 tokens on Free, 5,000 on Starter, 15,000 on Advanced and 30,000 on Pro. The pricing section explains how those figures look inside the app today, model by model.

A reviewer found strong long-context coherence, but after about 50 turns a secondary character's name was forgotten. That matches the pattern: the main line holds longer than the side characters, so the supporting cast deserves a line each in the notes.

I'd score memory as fair rather than strong. It is good inside a scene and needs help across a book, and that help costs either a bigger plan or your own discipline in the scenario notes.

The image generator

What does exist is an image generator, and a third-party review notes that it runs separately from the chat: "There's also a separate image generator that runs independently of the chat". In practice it is a side room: you make pictures there and carry them into your story yourself.

The "Save Your Images" modal is not decoration. DreamGen does not store generated images, so they have to be downloaded: "We do not store any images. Make sure to download any images you like." Image prompts are stored locally on the user's device: "Prompts are stored locally on your device."

I'd treat that like footage on a card with no backup. If you like a picture, download it before you close the tab, because nobody else is keeping a copy.

What it costs and what it allows

In the 23 September 2026 check, one safe-for-work image was generated with Z-Image Turbo for 1.69 credits in about 20 s. Desk research noted that the image credit cost is shown before generating but not published anywhere else. The cost section lists the price of every image model.

The image models are SFW-only. News of Sep 10, 2026 announced the new image models GPT Image 2.5 Flare and Sunburst.

My cover-art Saturday lives in the cost section, because in the end it was a story about credits more than pictures. The short version is that moody book-cover work is still allowed under the current rules.

If voice, calls or video matter to you, this is not your app. If you want words with the odd picture to set the mood, the generator does that job and asks you to keep your own files.

What free actually includes

The free tier's recurring allowance, under Account > Usage, is 200 monthly credits plus 25 daily credits. The daily credits only kick in after the monthly credits run out. The pricing FAQ describes them the same way: "a small daily usage allowance that only activates after you've used your monthly usage limit."

Recurring credits don't roll over. The monthly reset is tied to the account creation date, not to the calendar. I'd note that date somewhere, because it is the day your meter refills, and nothing else in the app will remind you.

Free gives access to all current models, paid with credits. Free has no credit packs. So the free plan is a sampler: every dish is on the table, and the plate is small.

Where free stops

The bigger limit is not credits but context. Free has a 5,000-token context on first-party models. Third-party models allow more: DeepSeek V4 Pro showed "Max input tokens: 30,000" and DeepSeek Flash 15,000.

There is a catch attached. Those larger-context third-party models carry a "may use your data for training" label. I'd rather write a short story with a small memory than hand a manuscript to a provider that may train on it.

For long fiction, I think the free context is too small to judge the product. A scenario with a proper codex eats a good part of the window before the story even starts.

A reviewer found that the free tier runs out quickly, in an afternoon of use. Both can be true. A short check barely dents the pot, and a writer with long prompts on the premium models can empty it in one sitting.

My view: the free tier is honest and useful for learning the tools, and nothing here nags you to pay. It is a test drive around the parking lot, though, not a road trip. If the scenario editor makes sense to you on free, that is the signal to pay for context, not for more credits.

Who free is for

Free is enough to learn the scenario editor, read a few public scenarios and find out whether steering suits you. It is not enough to judge memory, because the story never gets long enough to forget anything.

If you stay on free, treat the daily top-up as a safety net rather than a plan. By the FAQ's own description it is a small allowance, and it only starts once the monthly pot is empty.

The first upgrade was curiosity. The second was a practical decision, the kind I make when a client's footage won't fit on the drive I brought. Ada's town needed more room than the smaller plans gave it, and paying for the bigger window was the only way to keep the fog and the ferry in the same frame.

The prices have barely moved. I paid about $7.83 for Starter in 2024, and Starter costs $7.83 a month today. I paid around $19 for Advanced; the same plan costs $19.35 today.

Free and Starter on 23 September 2026

Free includes 200 monthly and 25 daily credits. Starter costs $7.83 a month, and its first month costs $6.26 (-20%). It has a 5,000-token context.

Starter includes 375 monthly credits plus 45 extra daily credits. It gets 0% discount on credit packs and on third-party models. In 2024 Starter was my escape from the free context; today Free and Starter share the same 5,000-token window, so the move I made in week 1 would buy credits and not memory.

Advanced

Advanced costs $19.35 a month, and its first month costs $15.48 (-20%). It has a 15,000-token context on Gemma 4 26B A4B and Qwen 3.5 27B. On GLM 5.2, GLM 4.7 and Gemma 4 31B it has an 8,000-token context.

Pro

Pro costs $48.30 a month, and its first month costs $33.81 (-30%). It has a 30,000-token context on GLM 5.2 / 4.7 and Qwen, and a 60,000-token context on Gemma 4 31B and Gemma 4 26B.

Pro includes 2,250 monthly plus 75 daily credits. It gives unlimited use of all first-party text models. It adds a 20% discount on credit packs. It also advertises "Up to 60% discount on select third-party AI models".

What "unlimited" means

"Unlimited" means 0 credits per use, under a fair-use policy. The pricing FAQ puts it plainly: "Unlimited means you can use specific models without worrying about credits - the credit cost for those models becomes zero." The fair-use policy rules out account sharing, bots or scripts, and excessive load.

The wording around it is less tidy. Desk research of the pricing page listed "unlimited basic and premium AI models" for Starter, Advanced and Pro. The notes from the 23 September check record the Starter card as offering "Unlimited access to basic/premium AI models", while the FAQ says "Free & Starter: Pay credits for all model usage".

I'd trust the FAQ and the crossed-out rows over any headline, and read every asterisk as a cost.

Here is the money math the way I'd do it. Advanced at $19.35 buys free use of the small models, with a 15,000-token window on them. Pro at $48.30 buys free use of every first-party text model and a window of up to 60,000 tokens.

That is well over twice the price, and it only pays off if the better models are the ones you write with every night.

My advice: skip Starter unless you only want more credits, because it adds no memory over free. Advanced is the sensible writer's plan if you are happy on the small models, where both the meter and the wider window work in your favour. Pro is for people who will write most nights, and at that point the price is closer to a gym membership than a snack.

My cover-art Saturday

It started as a break from writing and turned into a whole grading session. Each cover was a little moodier than the last, and each one cost credits I wasn't counting. By the time I looked at the balance, the month's allowance was gone and the month wasn't.

The lesson was not that pictures are expensive. It was that the words and the pictures drink from the same pot, so every cover I made was a few pages I couldn't write that month.

What a credit buys today

In the 23 September 2026 check, 2 role-play replies on GLM 5.2 cost 1.63 credits (2184 input and 906 output tokens). One Z-Image Turbo image cost 1.69 credits. By that rate, 200 credits is very roughly 200+ premium-model replies at small context, or about 100 cheap images.

That math is the whole trap in one line. A reply and a picture cost about the same, and a picture feels like nothing while a reply feels like work.

The model picker gives estimates too. For 10k input and 100 output tokens it estimates 3.43 credits on GLM 5.3, GLM 5.2, GLM 4.7 and DeepSeek V4 Pro. The same estimate is 1.72 credits on Gemma 4 31B and DeepSeek V4(.1) Flash.

On Gemma 4 26B A4B it is 0.34 credits.

Image prices

At default settings a Z-Image Turbo image costs 1.69 credits. A GPT Image 2.5 Flare image costs 3.54 credits.

A GPT Image 2.5 Sunburst image costs 3.54 credits too. A GPT Image 1.5 image, now deprecated, costs 6.75 credits. A Nano Banana 2 image costs 23.63 credits.

Packs and the meter

Credit packs are for subscribers only. Pack prices are not visible to free users; the page shows "Subscribe To Buy Credit Packs".

Pack credits are used after the recurring credits run out. They don't expire with the monthly reset.

The meter itself is transparent. The Usage page breaks usage down per day by model and token count. It also has a credit calculator. An adjustable input-token cap helps save credits.

My view: DreamGen's credit system is fair and well documented, and it still needs watching. I'd check the Usage page once a week, the way most people check a bank app, and keep pictures for the moments that need one.

There is no daily check-in reward. No referral or invite programme was seen in the app.

After sign-up only the verification e-mail arrived; no welcome or coupon e-mail came within about 30 min. I count that as a point in its favour. Apps that shout discounts at you in the first hour usually want you to stop reading the price.

The one real discount

Every paid plan has a first-month discount, shown on the landing page and the billing page. Starter and Advanced show "20% off for 1 month". Pro shows "30% off for 1 month".

No promo-code field was seen before checkout; the checkout itself was not opened. So the first-month price is the deal, and I'd skip hunting for codes.

I think the discount is fair and a little misleading at the same time. It is a real saving, but it is one month. The second month is the price you are actually signing up for, so judge the plan by that figure.

Earlier contests (#tldr2026, CYOA) had the same prizes. The contests say "every eligible creator gets a reward". For a writer that is the best deal on the site: build something original, publish it, and the app pays you in its own currency.

The catch is the word original. Fan-fiction fills most of the public shelf, so the contest nudges creators toward their own worlds, and I think that is a good nudge for a writing tool.

Promos in the news feed

The news feed occasionally announces free-model promos, for example GLM 5.3 Flash, "free through the end of August". Those are worth a glance if you already check the feed.

My advice: take the first-month discount if you were going to pay anyway, and enter a contest if you write original scenarios. Otherwise, pay the list price and spend the effort on the story.

What is allowed by default

Adult content in text is allowed by default for adult users. Only illegal content and sexual content that breaks the age rule are banned. That is a short list, and a third-party review reads it correctly: the rules mean "freedom for adult creative writing between consenting fictional characters, not a free pass for anything."

The filters you can switch

Account > Beta Settings > Text Content Moderation has a "Required moderation" filter. It blocks content that breaks the age rule and is "already enforced by system defaults". You can't switch that one off, and that is how it should be.

An optional toggle, "Sexual content moderation", labelled "Filter sexual content", works as an opt-in filter for anyone who wants a cleaner session. Beta Settings also has a "Moderation Sandbox".

Behind the switches, automated moderation scans content. The terms-and-privacy FAQ says the same: "DreamGen uses automated systems to enforce its Terms of Service and safety rules".

Images are stricter

All 5 image models are labelled "This model is only suitable for SFW content". Image generation is effectively SFW-only now. I'm fine with that split: words get the freedom, pictures get the brakes.

The age check

The age rule is a self-declaration, and no ID check is described: "By using the Service, you represent and warrant that you meet this age requirement". The Profile page states that an age check is not required in the viewer's location, so the check depends on the region.

The persona template seen in the check asked for an age field marked adults only. I'd call the whole gate honest but light: it records a promise rather than checking one. For a writing tool used at a desk that is the norm, and I'd rather see it stated plainly than dressed up.

The public shelf

The public scenario list surfaces many anime and fan-fiction self-insert settings from franchises whose main casts are written below the age of majority. Those cards were not opened.

The same shelf also carries a separate set of cards with grown-up themes, so the mix is wide. I'd like the franchise settings and the grown-up cards kept further apart in browsing, and I think the platform's own rules argue for the same thing.

Swiss law governs the Terms, with the CISG excluded. Disputes start with a 30-day informal resolution via [email protected].

After that comes binding arbitration under the Swiss Rules: seat Zurich, a single arbitrator, in English. The Terms also include a class-action waiver.

In plain terms: if something goes wrong, you e-mail first and argue in Zurich later, alone. For a subscription this size, that means the e-mail is your real remedy.

Who may use it and what is banned

Users must be "at least 18 years old and of the age of majority in your jurisdiction." The Terms ban anything illegal in the US, the EFTA countries or the user's own jurisdiction.

They also ban sexual content that breaks the age rule, and they spell out that fictional characters count. Beyond that short list the rules are very permissive. The About page stresses creative freedom for private content.

Your content

The Terms say the service claims no ownership or copyright over user content: "The Service does not claim ownership or copyright over any User Content". For a writer that is the clause that matters most, and it is the right one.

The other half is a broad licence. Users grant DreamGen an "unrestricted, worldwide, transferable, sublicensable, royalty-free license ... for the purposes of operating, providing and improving the Service." It is standard language, but it reads large, so I'd keep the finished manuscript in my own files.

Refunds

The Terms say all purchases are final and non-refundable, except where the law requires or at the service's sole discretion. The billing FAQ is more generous: "We offer prorated refunds". Refunds are requested by e-mail to [email protected].

Training and who reads your chats

DreamGen does not train AI models on user data unless the user opts in, and the opt-in is off by default. Data from users who did opt in is not removed from existing training sets. So opting in is a one-way door for anything already used.

First-party models run on DreamGen's own infrastructure. Requests may go to third-party providers under high demand or for certain features.

Some third-party models in the picker carry the label "Privacy notice: This model provider may use your data for training", for example DeepSeek V4 Pro and V4 Flash. The "Muse Spark ... Contributor" variants state that "data sent to the model may be used by Meta for training".

Private chats are not read by people unless they are published or submitted for publication, the user consents (for example in a moderation appeal), the user opts into model improvement, or the law requires it. That list is clear, and I think it is one of the better privacy stances in this category.

A cockpit, not a car

The UI is fairly dense and power-user oriented: tokens, credits, many knobs. I like knobs, and I still think a newcomer will stare at the session settings like a client at a grading panel.

Several stacked modals appear on first use: the Terms and age gate, the fiction disclaimer, and 3 image-generator modals. Each one is reasonable. Together they feel like signing for a parcel five times.

"Fox and fern books" had no opening, so the user must write first. A first-time visitor may stare at an empty box wondering what they broke.

Images are not stored server-side; they must be downloaded.

The pricing copy is flagged as inconsistent as well: per the check notes, the plan cards say "Unlimited* access to premium AI models" for Starter and Advanced, but the FAQ says Starter pays credits for everything and Advanced gets only the small models unlimited.

The Terms say purchases are non-refundable, while the billing FAQ promises prorated refunds. The company identity is shown only as an image, not as text. None of these is a scandal, but together they read like an edit nobody watched through before export.

API keys can still be created in Profile. Under Profile, "API Keys" lets any account create keys. So the door is open and the manual is missing.

Finding the company

DreamGen goes by several names: "dreamgen ai", "DreamGen V2" and "v2.dreamgen.com". The blog lives at dreamgen.com/blog, with an RSS feed. Support is by e-mail at [email protected].

Discord is the community hub, with an invite reached via dreamgen.com/discord. It is the only social link in the site and app menus.

The landing page footer has no social links. No X, Reddit, Instagram, TikTok or YouTube links were found.

The Hugging Face page at huggingface.co/dreamgen hosts open-weights 'Opus' story and role-play models, such as opus-v1.4-70b-llama3, opus-v1-34b and opus-v1.2-7b. It was found by web search; the app does not link to it. For a company proud of its models, that is a strange thing to hide.

No phone app

No official mobile app was found. No mobile app-store links were found either. An iTunes search for "dreamgen" returns no DreamGen app, and the Play id com.dreamgen.app returns 404.

For a desk-bound writing tool that is a small gap, but I'd be wary of any store app that claims the DreamGen name.

It suits users who want API access or open-weights models. It also suits adults who want permissive text content with clear privacy terms.

Who should skip it

It is not for people who want a simple "AI girlfriend/boyfriend" companion with voice calls, selfies or a mobile app. Nobody here will ask about your day.

It is not for users put off by credits and tokens jargon. If "max input tokens" makes your eyes glaze, the first evening will feel longer than a wedding reception.

It is not for anyone who wants free unlimited chat either: the free tier is 200 credits a month with a 5K context. That is a test drive, not a home.

The honest summary is that DreamGen rewards effort. The people who get the most from it are the ones who would happily read its docs, and the ones who bounce are usually right to.

My rule of thumb: if you have a story you want to finish, try it. If you want company, look elsewhere.

Cancelling runs through Account > Billing, then "Cancel Subscription" at the top of Manage Subscription: "Click "Cancel Subscription" at the top". It is two clicks, not a maze, and I give the app credit for that.

Subscriptions auto-renew and are billed "in advance on a recurring basis". So the charge lands before the month you are paying for, which is normal and still worth a calendar note.

Changing plans

Plan changes take effect immediately. Upgrades are prorated. A downgrade gives an account balance, not a refund.

The FAQ is clear on that last point: "The prorated difference is credited to your account balance ... No refund is issued to your payment method". I'd think twice before an upgrade you might regret, because money that goes up comes back down only as store credit.

Deleting your account

Profile has a "Request Account Deletion" button. After deletion, data is removed from active systems "within a reasonable period". Backups keep the data for a while longer after deletion.

The terms-and-privacy FAQ gives a number for that period: deleted data leaves active systems typically within 35 days ("typically ≤35 days"). I'd export or copy anything you wrote before pressing the button, because the button does what it says.

Before you pay

My checklist is short. Learn the scenario editor on free before you pay for context. Note the date your plan renews.

If you only want a month of writing, pay for one month, write hard, and cancel in the last week. The access runs to the end either way.

One more habit I'd keep: copy your best pages out of the app every week, not only at the end. A paused plan or a deleted account should never be the moment you discover which chapter you forgot to save.

Tune the meter

Lowering max input tokens saves credits. I'd set it as low as the story allows and raise it only for scenes that need the long memory. It works like shooting at a lower bitrate for B-roll: nobody notices, and the drive lasts longer.

Keep the scenario notes tight for the same reason. Every line of lore rides along with every request, so a bloated codex is a tax on every reply.

Read the manual first

The docs cover Role-Play, Story, Scenario Editor, Wizard, CYOA, Publishing, Image generation, V1-to-V2 migration and community guides (Google Docs, rentry). I'd read the Story and Scenario Editor pages before the first evening, not after it. Half an hour with the docs saves the IKEA experience.

Spend like a writer

Pick the plan for context, not for credits. If the small models write well enough for your story, the plan that makes them free is the smart one; if you need the best models every night, pay for the plan that makes those free instead.

Keep pictures for covers and references, download every one you like, and don't let a mood board eat the month. Save your best pages outside the app as you go. Finally, write the steering notes like a director: short, specific and about feeling, not plot.

A note like "colder" does more than a paragraph of instructions, the same way one clear note to a colourist beats a page of adjectives.

Memory is mixed. Inside a scene it holds, and across a long story it drifts unless you keep the key facts in the scenario notes or pay for a wider window.

The free tier is mixed: honest, no ads, all models, but a context too small for long fiction. Value for money is mixed too, because the useful plans hide behind credits, asterisks and per-model windows.

Images and voice rate weak: no voice, no video, and pictures that are an extra rather than a feature. Privacy and terms rate well, thanks to opt-in training and clear model labels, even with the company name hidden in an image.

Cancelling and billing rate well, with an easy exit and a refund policy that is friendlier in the FAQ than in the Terms. Content settings rate well: permissive for text, strict for pictures, with a light age gate.

The apps in that table are companions first: a character, a relationship, a chat that asks how your day went. DreamGen is a story-writing and role-play engine rather than a companion app, and it says so. That one difference decides most of the comparison.

It wins on privacy posture as well. Opt-in training and labelled third-party models are rarer in this category than they should be.

Where DreamGen is worse

It loses on everything a companion is supposed to be. There is no voice, no calls, no phone app, and nobody remembers your birthday. The first evening is homework, where most companion apps have you chatting in minutes.

If you want someone to talk to, pick one of the others in the table. If you want something to write with, DreamGen is the only one of these I'd open.

More on this site: the best AI girlfriend apps, free AI chat apps, the best AI companion apps, AI roleplay chat apps, all AI companion app reviews.

It is not a polished product. The copy contradicts itself, the menus assume you already speak its language, and the prose sometimes sighs like a bad poet. None of that mattered much, because the core of it worked: a story took direction the way an edit does, and the story answered.

I also respect how it treats the writer. Your words stay yours, training is your choice, and the rules are permissive where fiction needs room and strict where it should be.

So my last word is simple. If you have a book in a drawer, DreamGen is a good way to get it onto the desk. Just don't expect it to write the book for you.

That part, annoyingly, is still the writer's job.