On May 8th, "History Simulator: Chongzhen" officially launched on Steam. As the "first AI-native historical strategy game," it attracted significant attention from players and media even during its testing phase, with many gameplay videos garnering high popularity.
However, after the game's official release date was set, public opinion took a sharp downturn, quickly plunging it into controversy.
The core issue lies in the game's payment model: a one-time USD 48 buy-to-play game package, combined with an in-game purchase system for Token (word unit) credits.
The buy-to-play package includes an initial computing power data allowance, sufficient for approximately 30 hours of gameplay.
Once this initial allowance is depleted, players must top up to acquire credits for Token consumption.
The consumption rate is roughly USD 1-2 per hour. The game also offers an optional advanced model (Expert Mode), with consumption rates varying from 2 to 8 times that of the basic mode (approximately USD 2-16 per hour).
The development team stated candidly in the launch announcement that they are "selling at cost, not aiming to lose money" and "not profiting from token price differences."
Rationally speaking, the developer's words ring true. Previous promotional materials from Zhipu disclosed that the Token consumption for a single complete session of "Chongzhen" reaches tens of millions [1].
Calculated using the basic model, DeepSeekV4 flash, the per-usage API cost exceeds USD 5.
For Expert Mode, which uses GLM-5 Turbo and Gemini 3, a rough estimate based on official website pricing puts the consumption cost at over USD 15 per session.
But this reasoning clearly doesn't sit well with the player base.
Steam's first-day reviews were "Mixed," and the comments section below the promotional video was filled with negative voices and skepticism, with most players directly accusing it of being overpriced and having an unreasonable charging method.
Player dissatisfaction reflects deeper, structural contradictions within AI-native games.
The business logic of traditional games is built on a marginal cost approaching zero. Once a game is developed, the server cost of attracting one more player is negligible.
After breaking even on development costs, any new player and any payment they make is almost pure profit.
But AI-native games shatter this logic.
Every new player entering the game means a tangible computational overhead. The larger the player base, the higher the cost—a cost structure that threatens to spiral out of control at any moment.
More troublingly, players have almost no intuitive perception of Token consumption. Unlike Manus or other Agent-type products, where a consensus on high token usage has already formed among users,
a simple instruction of a few dozen words from the front end might require tens of thousands of tokens behind the scenes for system prompts, context, and multi-step reasoning.
But players easily lack a concept or psychological expectation of how much computing power a session burns. This makes the pricing itself harder to understand and accept.
For instance, some in the comment sections quickly questioned why not just do text simulation with Doubao, or why not set up your own API with a custom prompt to play.
Even more critically, games like "History Simulator: Chongzhen" require a high level of model quality.
As a strategy game with an extremely long context window, a single session can span several hours. The model needs to stably maintain memory, run numerical calculations, and preserve the internal consistency of its world-building.
Much user feedback indicates that the current experience is still highly dependent on the quality of the underlying model, and the experience with the expert mode's Gemini 3 Pro is noticeably better.
If any inference goes awry, resulting in incoherent or contradictory narrative, the player's immersion collapses instantly.
To prevent the game's fundamental reputation from crumbling, the developer has to grit their teeth and use good models.
It's hard to compress costs by switching to cheap, smaller models or on-device models. At least at this stage, not a single cent can be saved on model costs.
On the other hand, from the player's perspective, a buy-to-play model should not become a disguised form of "rental."
In a buy-to-play game, the principle that paying a fee grants clearly bounded, complete gameplay rights is a long-standing, tacit understanding between developers and players.
No matter how legitimate the reason, technology or experience should not deprive players of their right to choose and their subsequent freedom to play.
Since they chose to make a buy-to-play game, I believe "Chongzhen" should hold that line. The initial purchase price should cover a quota sufficient for a user's first completion, ensuring the initial experience is uncompromised. Then, at the very least, allow users, after their first playthrough, to plug in their own API key and choose their own payment model as needed.
The development team could also try to visualize Token consumption more clearly, letting players see in real-time what they are consuming, how much they are consuming, and how many tokens the system prompts are taking up.
While it wouldn't reduce the cost of playing, it would enhance transparency and trust, helping to avoid much of the criticism labeling them as "middlemen scalping."
Beyond these measures, what the development team can realistically do is indeed limited. The real solution is likely still a waiting game—waiting for computing power costs to drop by another order of magnitude, waiting for more cost-effective, high-quality models to become widespread. Perhaps then, all current predicaments will become moot points.
Getting the product out the door is just the first step. Finding a more rational, sustainable business model is the true challenge AI-native games must overcome, and this will clearly require more exploration and trial and error.
References
[1] GLM-5 Drives the First AI-Native Game "History Simulator: Chongzhen", Zhipu