Since the first Large Language Models have hit the market, new fantasies of automatic content generation have lit up the faces of many executives and creatives in the gaming industry. Ideas of dynamic npc dialogues, automatic quest systems all the way to dynamically generated games have made the rounds and yet, very little of this seems feasible. The idea of generative content is in itself not new and by the 1990s had already been well established in games like Elder Scrolls II. Newer systems include the Nemesis Systems of the Shadow of Mordor series. The same also goes for Chat Bots which have been around even longer. The promise of generative content is seductive, comparably little effort with tons of different outcome, keeping players in the games for much longer than hand designed quests could ever feasibly do. The problem is, that even with our newest models the numbers just do not add up.
Digital games beat physical media on marginal cost. Per-action AI inference gives that cost back. On the numbers published for Status, it gives back several times what the game earns per player.
A few weeks ago, Naavik published an article on AI native features, meaning features that could not exist without AI. Think of games like Teaching My Little Sister How To Drive. While some of the features seem more feasible then others, these do not really sound like million dollar ideas to me and are easily copied. An issue that many game developers have already lamented about in the recent Steam Next Fest. My interest here mostly the game Status, which comes with a real vision of a big AI Native game. It might not be an original one but definitely of a big feature. As the article describes it:
“At its core, Status is a simulated social network in which a player creates a persona and interacts with a familiar social feed by posting, replying, and sending DMs. The difference is that the “people” responding are AI-driven characters with their own personalities, relationships, and memories.”
And with this game description also come some numbers that we can create ourselves an actual business case.
Digital products, which modern games largely are, have a huge advantage to physical media. One that silicon valley has been exploiting again and again, be that in Google's rise or Facebook's or any other of the magnificent 7. Offering digital products means that you can offer an infinite amount of your product at 0 variable costs. Meaning it costs us almost nothing in addition to the fixed costs to sell to 1 user or 2 users.
Introducing per player AI requests to a game is breaking this dynamic. Now we create per player costs per day they are playing, which creates a much tighter ROI corridor to hit for your Marketing team, since the ARPDAU is already coming with a direct subtractor.
As described above Status by WishRoll is a simulated social network. The subtitle is even "Sims but Social Media". Players can post on their own little network and AI characters answer, creating a deep immersion, that is hard to achieve otherwise, and even the best by hand designed systems are easily breakable. During their Beta, the Studio reports costs per player of 12$ to 15$ per day. They have claimed to have reduced theses costs by about 95% which would at a generous spot leave us at 0.60$ per user per day. That daily user base is supposed to be made up of 500K DAUs at 1.5 hours of daily playtime. At first glance that sounds pretty good. Significant cost cuts combined with a pretty big user base. They have also raised 17M$ across Seed and Series A as announced in May 2026.
The game monetizes it self quite traditionally through purchases, subscriptions and rewarded ads, which as per the article seemed to be quite successful.
Lets apply these numbers for a real business case. The article shows the Sensor Tower Statistics on ARPDAU and D1 and D14 retention. Sensor Tower data is only an estimation and can be somewhat off, but I am doubting that it is off by a lot. From 2025 to July 2026 through changes in monetization rose from a few cents to around 0.20$. These changes mostly concentrated on player actions and introduced systems like an energy system to regulate the amount of AI calls a player can trigger. But combine that with the reported 500K daily actives that would leave us with daily revenue of about 100K$. The issue here is, that we also have to make the same calculation for the AI costs, which as we estimated earlier lie around the 60 cent marker. Which means the daily AI costs based on these numbers amount to 300K$ per day. That leaves us with a difference of around 200K$ which seems unlikely or unsustainable.
Additionally their interesting and immersive game design seems to work against the product. The quoted retention numbers are very good compared to the rest of the mobile market. Naaviks article shows Day 1 retention values of 50 to 55% and Day 14 values of 13 to 15% at first dropping down to 43% for D1 and 8% by D14. Still, these numbers might make you want to build your one Status game but remember the costs are scaling with the amount of players you have. So high retention numbers are working against us here. The best case for a game with these costs per player is that they register, pay, and never come back, assuming that we did not spend much on acquisition. Otherwise it might be more dire.
Now, not everything about AI native features is doomed from the beginning on a financial level. Another reported game is Death by AI by Little Umbrella Games. This is a social game similar to games like Cards Against Humanity where everyone in the group has to convince an AI Bot not to kill their avatar.
"Death by AI is a prompt-based game with a core mechanic that is deceptively simple. Players are dropped into a deadly scenario (like being cornered by a venomous snake) and given 60 seconds to explain how they intend to survive. There are no predefined actions to choose from; instead, players have total freedom to write exactly how they plan to escape. They can type whatever they want — whether ridiculous, clever, or elaborate — after which an AI Game Master interprets the response and decides whether the player lives or dies. The AI Game Master also graciously explains exactly how their fate unfolds."
The game had huge success, launching through Discord but on the Stack of Open AI plus Eleven Labs voice generation, had similar issues as described above and become unsustainable in three days. They managed however to move to smaller models, which the scope of this game should easily allow form and maybe new decision models like JEV will make this even cheaper. JEV is a model that picks an action from a list of legal actions for a given game state, and does not generate free text (Seeles explainer). The cost saving is my own speculation, as the sources give no costs for it.
Another Game is MIMESIS which is a coop horror game, where the game records your voice during the game and an AI model tries to imitate you to your friends by choosing voice lines from the recorded material. This will also be hugely helped with this new era of decision models. The model here sits presumably locally, and seems to be rather small and extra trained for this task. So the 9.99$ price point creates direct revenue with no new cost per player.
And also Status has analysed the problem correctly and moved to introducing better monetization that limits the players action with them paying for it. This is however more a patch over the issue than a real solution. Monetizing and locking features should work to increase the possible play experience and therefore monetizing players who are willing to pay for it, not compensating for a cost leak in the basic system of a game.
Before shipping a per-action AI feature:
So when conceptualizing a new AI per action feature you should be able to answer a simple question with yes: Can your Monetization compensate for the new variable costs you are introducing?
This may sound like a trivial question, but not all player bases are built equal. Not all player bases take as well to introducing new features with incredibly hard pay walls especially in the live-ops genre. So think about the opportunity costs that could also come from unhappy customers. Another question is, what activity behaviour are we incentivizing for ourselves with this feature? If we would want low activity with high pay activity, that might not be a feature built for the long term and our time could be used with other features.
The verdict in this case is, it's the AI that kills you. These costs if correctly reported, are unsustainable basically from day one– even with cost reduction through better deals with vendors. There might be more to it in terms of investment or revenue sources that this article is not reporting on. Another mitigating idea could be to either set up your own AI servers, but it is also more of a cost optimizing measure than a real solution. The best solution would be to ship small language models in the game. HOWEVER: this is inviting players to jailbreak your models, using it for purposes they were not intended for, and could become a compliance or PR desaster.