AI Costs Hit ¥160,000/Day, Even a Bank Loan… “Teaching My Sister to Drive” Devs on Demo’s Explosive Popularity [Interview]

"Teach My Little Sister How to Drive" is a driving instruction simulation currently in development for PC (Steam). Players give verbal instructions from the passenger seat to teach their "sister, who just got her license"—controlled by AI—how to drive.

The sister doesn’t always follow the player’s instructions; sometimes she acts unexpectedly or even talks back. The unexpected moments that arise from these free-form voice conversations have garnered attention on social media and video streaming platforms, and the game has become popular among many players in Japan as well.

However, this popularity has placed an unexpected burden on the development team. As Game*Spark reported recently, the number of players using the demo version increased approximately 20-fold in just one month.AI service fees have reached over $1,000 (approximately 160,000 yen) per day, and usage restrictions on the main AI model are causing issues with in-game responses. The developer, Easy Fox, has revealed that it is securing operating costs through a bank loan and that it may end the demo version earlier than planned.

Developer Cries Foul Over AI Costs Exceeding 160,000 Yen Per Day. “Teaching My Little Sister to Drive” Player Count Rises 20-Fold in One Month—Early End to Demo Version Also Under Consideration | Game*Spark - Domestic and International Gaming News Site

It has also been stated that the expected token usage fees for the full version will be included in the game’s price and will not be billed separately.

https://www.gamespark.jp/article/2026/10/07/173304.htmlRead more »

So, Game*Spark conducted an email interview with the developer, Easy Fox. We spoke with them in detail about everything from why the AI service issues haven’t been resolved to the circumstances that led them to secure a bank loan, the possibility of offline play using local AI, and their outlook leading up to the official release in January 2027.

“If you want to use it more, you have to use it more”—The dilemma created by Gemini’s usage limits

――This game uses Google’s “Gemini 2.5 Flash” as its primary AI and leverages OpenAI’s Realtime model as an alternative service. Please explain why Gemini is well-suited for this game and the differences resulting from switching to the alternative model.

Easy Fox: Of all the models we’ve tested so far, Gemini 2.5 Flash proved to be the best fit for this game. The main reasons are its fast response speed and its ability to express the character’s personality.In this game, it’s crucial that the AI doesn’t just understand and execute commands, but reacts like a real younger sister and engages in fun conversations with the player. Gemini 2.5 Flash delivered the most satisfying results in that regard.

On the other hand, the alternative model we’re currently using has issues such as conversations becoming less engaging than before, or the sister talking for longer than necessary. In this game, conversations with the sister aren’t just an added feature—they’re a core element of the game’s appeal. Therefore, we believe these differences have a significant impact on the player experience.

――I understand that raising the usage limit for Gemini requires meeting certain usage criteria. What makes this issue particularly challenging? Is there any prospect of a solution?

Easy Fox: The biggest challenge right now is that solving this problem isn’t something we can achieve through our own efforts alone—it depends on Google Cloud taking action.

To raise the usage limit for the Gemini model, we need to upgrade our Google Cloud service tier. However, we were informed by the support team that they cannot manually raise the tier; instead, we must meet the upgrade criteria by using more of their services.

The problem is precisely that our usage is already capped. To raise the usage limit, we need to use more of the service. However, because of the current usage limit, we cannot use it beyond that point. This puts us in a very difficult, effectively dead-end situation.

We are still inquiring to see if there are any other solutions. We sincerely hope that this article reaches the relevant personnel at Google Cloud and that they will work with us directly to find a solution. At the same time, we are proceeding with technical preparations in parallel so that we can migrate to other AI services.

[Editor’s Note] The information regarding Google Cloud usage tiers and limit increases is based on explanations provided to the developer by Google Cloud representatives.

――This title features a system that automatically switches between multiple AI providers. When the model changes, what is the most challenging aspect of maintaining consistency in the dialogue and character portrayals?

Easy Fox: The biggest challenge is that each AI model has its own way of role-playing and interpreting characters. Even when given the same character profile and conversation prompts, not all models will portray a younger sister with the same personality.

One model might respond with excessive kindness, while another might stick to purely explanatory conversations. There are also models that cannot fully capture the mischievous and slightly mean-spirited personality of the younger sister we intended. Ultimately, we have to create and fine-tune individual character scripts for each model to ensure it accurately recreates the desired personality.

An even greater challenge lies in language and cultural differences. This game currently supports 10 languages. Therefore, we need to fine-tune the younger sister’s personality and conversational style based on the combination of each language and AI model. However, it’s impossible for the development team to perfectly understand the languages and cultures of every country.

Conversations between siblings that feel natural in Japan may come across as unnatural in other cultural contexts. We want to create a younger sister who feels much more natural and truly like a real younger sister. To that end, we are currently gathering feedback on character personalities and dialogue from players in various countries.

We also have a request for everyone reading this article. Please join our Discord and let us know what kind of personality you envision for a “little sister,” and what kind of conversation you would have with her if you were actually teaching her to drive—it would be a huge help.

Local AI Development Is Complete—But Why Is Going Fully Offline Still Challenging?

――You’ve announced plans to introduce a supplementary local AI model for players with high-performance GPUs. Could you tell us about the current development status and the remaining technical challenges?

Easy Fox: Actually, development of the auxiliary local AI model itself is already complete. Our original plan was for the local AI to work in conjunction with Gemini.

No matter how fast Gemini is, there are times when its response is delayed. However, since driving requires quick decisions and reactions, there’s a risk that my sister might be unable to drive properly while waiting for the AI’s response.

We designed the local auxiliary model to compensate for those moments. In other words, while the cloud AI handles natural conversation and complex decision-making, the local AI plays a supporting role in situations requiring quick reactions.

However, a new, unforeseen problem has arisen: it has become difficult to use Gemini, our main model, reliably. As a result, we are testing how useful the already-completed auxiliary model is in this situation and how we can resolve these new technical issues.

――Local voice generation technology is already being used in this game. On the other hand, why is it difficult to run the entire dialogue system on the player’s PC?

Easy Fox: Actually, we’ve received feedback from many players via X and Discord stating that “not implementing a local LLM is an inefficient choice.” We ourselves are very interested in the potential of local LLMs and have already tested various models.

However, given the current state of technology, we’ve determined that an LLM of at least 8B (8 billion parameters) is necessary to achieve the level of natural and engaging role-play we’re aiming for. The issue lies in the hardware requirements.

In the runtime environment we envision, a GPU with at least 8GB of VRAM is required to run an 8B-class model smoothly. Furthermore, if we run a voice generation model simultaneously, we believe that, in practice, at least 12GB of VRAM is necessary for stable operation. This represents a high hardware requirement for a significant number of Steam players.

While local speech generation performs relatively limited functions, conversational AI must understand the player’s words, assess the situation, generate responses that match the character’s personality, and even determine driving behavior. Consequently, there is a significant difference in the computational power and model complexity required for each.

We have tested various small local LLMs, but at this point, we have determined that it is difficult to achieve the desired level of “little sister” behavior using only small models. Of course, we acknowledge that our assessment may be incorrect.

If anyone knows of a multimodal model in the 2B to 4B range that supports direct voice input, achieves an average response time of about one second in a typical user environment, and delivers excellent role-playing performance, please do contact us. We’d be happy to learn more.

――In the long term, do you think fully offline play using local AI will become a reality? Or will cloud AI continue to play a central role?

Easy Fox: Considering the pace of AI technology development, I expect that by the end of 2027 or around 2028, compact local AI models with performance levels we find satisfactory will emerge. Of course, this is merely my personal prediction.

Even now, I believe it’s entirely possible to create fairly high-quality local AI games for users with GPUs equipped with 8GB of VRAM—provided real-time voice dialogue isn’t required.

However, from a game developer’s perspective, there is one troubling issue. When creating games, we naturally want to achieve richer dialogue, more complex situational judgment, and more convincing characters. Doing so requires higher-performance AI models, which means we’ll likely end up relying on cloud services once again.

We feel the same way. While we look forward to a future where fully offline play is possible, what matters most is how entertaining and polished an experience we can provide to players. Therefore, we intend to continue weighing the pros and cons of local AI and cloud AI, choosing the path that maximizes game quality.

“Let’s not go into debt just to make a game”—Maintaining the demo version even at the cost of breaking our initial promise

――It has been revealed that operating the demo version currently incurs over $1,000 per day in AI service fees, and that you even took out a bank loan. In light of the unexpected surge in players, what was the most difficult decision you had to make?

Easy Fox: First of all, I felt truly grateful. Actually, my team members and I had joked about this before:

“What if, even before we’ve released the game, so many people start playing that our server costs become unmanageably high?”

And we replied, “If that ever happens, we’ll pay the costs with gratitude.”

But that’s exactly what happened. What’s more, the period when player numbers exploded coincided with a long holiday weekend. Even while spending time with our families, we were quite shaken as we watched the costs skyrocket.

When we returned to the office after the holiday, we first decided—just as we’d discussed before—to cover the costs with a sense of gratitude. Of course, the reality of the situation was far more serious than we’d imagined.

As expectations for the game grew, we wanted to provide a great experience for even more players, but our main AI model became unstable, and our operating costs skyrocketed. Ultimately, we prioritized switching to an alternative model—even though it was more expensive and offered slightly lower conversation quality—so that many players could enjoy the game with relative stability.

As a result, I was unable to keep the promise I made to my team members when we first started developing the game. We had originally said to each other:

“At the very least, let’s not go into debt to make this game.”

However, the very next day after making that decision, we began looking for a bank loan and ended up taking one out.

Even now, I’m still not entirely convinced that continuing to shoulder such high costs was the right decision. However, there is one thing we believe: if we create a truly fun game, players will eventually recognize its value. For now, we’re continuing development, supported by that belief.

――We understand that in the official release, you do not plan to charge separately for AI usage, but rather include the estimated costs in the price of the game itself. If there are players who spend significantly more time than average conversing with the AI, how do you plan to ensure sustainable service operations?

Easy Fox: We’ve been running the demo version for about eight months and have accumulated data on player usage patterns and AI service costs. Based on that, we’re calculating the AI costs expected to be incurred by the average player while using the game, and we plan to set the package price for the official release version with those costs in mind.

Of course, not all players will play in an average manner. Some may converse with their little sister for much longer than we anticipate. In such cases, we’ll acknowledge that our cost projections weren’t sufficiently accurate, and we intend to absorb any resulting loss of profitability ourselves.

We do not want to create a game where players have to worry about additional charges every time they talk to their sister. We want to provide an experience that is as similar as possible to a typical one-time purchase game—where, once purchased, players can enjoy the game itself without worrying about costs.

――You’ve also mentioned the possibility of ending the demo period earlier than planned. Could you explain the criteria for making that decision and the impact this issue has had on your development schedule?

Easy Fox: The most direct factors are the rapid increase in the number of players and the resulting operational costs for the AI. We are a small game development team with a limited budget.

If costs continue to exceed our projections as they are now and we end up exhausting our entire budget, there is a possibility—in the worst-case scenario—that we may have to end the demo service early.

The biggest impact this situation has had on our development schedule is that it led us to decide we cannot delay the release date any further. We are aiming for a January 2027 release, and under the current circumstances, sticking to that schedule has become our top priority.

As a result, both I and the team members have recently been staying overnight at the office to continue development. Since we cannot increase the time allotted to us, we are focusing more intensely on development, even if it means sacrificing our personal time.

Of course, I don’t believe this way of working is sustainable in the long term. However, our most important goal right now is to seize the opportunity we’ve been given and deliver a game of sufficient quality by the scheduled release date.

“Intentional Game Design” Is What Upsets Players—But Technical Glitches Are a Different Story

――Issues such as unnatural dialogue, excessively long response times, and incorrect driving controls have been reported. How do you distinguish between “the younger sister’s mistakes,” which make the game fun, and “technical glitches,” which cause players discomfort?

Easy Fox: We believe there are two types of frustration players experience. One is the frustration caused by the younger sister being too silly or disobedient. The other is the frustration caused by the game system not functioning properly.

We’re aiming for the former.

In fact, the fact that the younger sister talks longer than necessary or doesn’t fully understand the player’s intentions—causing a sense of frustration—is, to a certain extent, an intentional part of the game design. In real life, too, when teaching someone to drive, there are moments when the other person doesn’t move the way you expect, and you feel frustrated. We wanted to transform those emotions into part of the game’s fun.

However, the unnatural conversations currently occurring are a separate issue. This is one of the problems arising from the fact that we can no longer reliably use our main model, Gemini.

Having your sister talk at length might be amusing in certain situations. However, if the conversation feels unnatural and loses its humor, it simply becomes an unpleasant experience for the player. AI responses that are too slow, uninteresting dialogue, or game system glitches that prevent normal progression—these are not the kinds of challenges we intended.

Our game is, indeed, designed to make players angry. However, that anger must stem solely from interactions with the character of the younger sister. It must not be caused by technical flaws or unreasonable game mechanics. We make a clear distinction between these two, and we are committed to resolving any issues that fall into the latter category.

――To what extent can the AI currently determine the younger sister’s actions on its own? Conversely, which aspects are controlled by the conventional game system?

Easy Fox: In the current demo version, the main elements directly handled by the AI are the younger sister’s driving behavior, facial expressions, and dialogue. The system is designed so that the AI determines her reactions and actions based on the player’s voice commands and the current situation.

Most of the stages included in the current demo version are designed so that the game can progress using just these three elements. However, the official release will feature even more complex driving situations.

For example, we plan to add stages involving driving on regular roads and highways. In these environments, simply determining how the younger sister drives isn’t enough. We’ll need a traffic system capable of handling various factors, such as other vehicles on the road, traffic conditions, and driving environments.

Therefore, in the official release, more traditional game systems will work in conjunction with the AI than they do now. Our goal is not to create a game where everything is left to the AI. Rather, we aim to combine the AI appropriately with existing game systems so that the AI can play the most interesting role.

Our goal is a “sister who truly feels alive”—not just one who goes viral in videos

—Looking ahead to the January 2027 release, what is your top priority right now? Also, please tell us about any aspects of this game that you’d especially like Japanese players to know about.

Easy Fox: When developing games, we always consider fun and an enjoyable gameplay experience to be the most important factors.

In fact, from the very beginning of the game’s planning phase, we aimed to create an experience that people would want to share through live streams. In that sense, we’re very happy that the game has become a hot topic and that many creators have praised it as a “game well-suited for streaming.”

However, our goal isn’t simply to create a game that goes viral. We believe that by leveraging AI, we can offer new and exciting gameplay experiences that were previously difficult to achieve.

Our ultimate goal is not just to have players choose from pre-set options like in traditional RPGs, but to let them engage in free-flowing conversation and truly feel as though they’re teaching their little sister how to drive.

As we prepare for the official release, our top priority is enhancing this sense of immersion. To that end, we’re working to resolve current technical issues and minimize any awkwardness caused by language and cultural differences. We’re striving to deliver a more natural and polished experience by the time of the official release.

I’d like to express my gratitude to all our players in Japan. So many of you have shown interest and sent us your warm support. We’ve also received some sharp criticism at times. We take all of these opinions to heart.

While there are still many areas where we fall short, we will do our absolute best until the very end to create a game that you will all find genuinely enjoyable. We want to present a younger sister who doesn’t just appear briefly as a funny character in a viral video, but one who feels like a living, breathing character throughout your entire gameplay experience.

I hope this interview serves not only to convey the operational challenges we’re facing, but also to provide an opportunity for you to understand the struggles and challenges of a small development team striving to create a new kind of game using AI.

The interest and support we’ve received from players in Japan have been a tremendous source of strength for us. We’d also like to express our heartfelt gratitude to all Game*Spark readers. We’ll continue to work hard to deliver a game that’s far more polished than it is now by the time of the official release.

――Thank you very much.

"Teaching My Little Sister to Drive" is scheduled for release on PC (Steam) in January 2027.