Pax Historia: First Impressions. An AI-powered grand strategy where victory depends on your prompting skills

Pax Historia — a browser-based sandbox strategy where the interface is replaced by a text field and the game engine by a language model. There are no familiar tax sliders or army caps here. Just you, a set of nations, and an AI on the other end of the line. No micromanagement in the classical sense: open the game, pick a scenario, and start reshaping history. Or don't. Let's break down what Grand Strategy becomes when a neural network writes it.

What if the Axis had won? What if the Cold War turned hot? What if the assassination in Sarajevo never happened? What if Rome held? Pax Historia is built on exactly these questions — and it doesn't just ask you to read an alternate history, it asks you to play one out. The developers, a YC-backed startup, pitch it as the first LLM-driven grand strategy: some players craft scenarios and publish them to the community, others pick a country and jump in. For an alpha, the numbers speak for themselves: 35,000 daily active users.

Genre-wise, it sits as close as you can get to global strategy games — but without their traditional skeleton. There are no formalized administration mechanics, demographic models, tech trees, or GDP calculated to two decimal places. That entire familiar layer is replaced by freeform text and the boundless imagination of a neural network.

The core loop is simple: word — action — reaction. The player states their nation's intentions, then initiates a time jump. At that moment, the AI simulates what happens, generates a narrative report of events, and updates the map.

This immediately raises an obvious question: what exactly is a turn in this game? Unlike EU4 or Civilization, where a turn is a fixed unit of time, in Pax Historia the player chooses the time span of each jump: anywhere from one week to a year, or until the next major event. On one hand, this offers flexibility. On the other, it blurs the sense of progress. It's hard to feel momentum when you're not entirely sure what happened in the gaps between your decisions.

One detail that guides consistently emphasize: the quality of your gameplay is directly tied to the quality of your prompts. Short, careless descriptions produce weak results — the model simply doesn't read the depth of your intent. In traditional strategy games you need to know the interface; here, you need to know how to express your thinking precisely and concisely.

Three core tools make up the main gameplay interface: actions, chat, and the advisor.

The game also already supports a choice of AI models (from lightweight to advanced), a turn rewind feature, and a full scenario editor.

Pax Historia runs on a token economy, and it's worth understanding the rules before your balance hits zero in the middle of a world war.

On registration you receive 1 token, and the daily bonus is 0.2 tokens (with a cap of 1.2 in reserve). Unspent tokens from the previous day expire. Tokens fuel everything: from generating a turn to every message sent to the advisor. Cost varies heavily depending on which AI "brain" you've selected. The Pro model (Gemini 3.0 Flash), for example, runs about 0.02 tokens per turn. In practice, one purchased dollar — equivalent to 1 token — burns through in roughly 15–20 minutes of active play.

There's a Pax Patron subscription that grants access to models through OpenRouter without per-token billing. For anyone planning to play regularly, this makes more sense than buying tokens piecemeal. The developers are upfront about it: they pay AI providers for every single request, and ad revenue won't cover that. For now it's the only way to keep the project afloat, though the team is actively looking for ways to bring costs down.

The project's main draw is absolute freedom. The ability to reshape the map and context lets you play out virtually anything: from "What if Napoleon won at Waterloo?" to "What if a zombie outbreak began in 1980?" The system imposes no "right path" — because no such thing exists in the code. There is only the AI's interpretation.

This shines brightest in diplomacy. You can construct elaborate political architectures and multi-layered agreements that the AI will genuinely attempt to honor. At that point it's less of a game and more of a "prompt strategy" — one where the depth of your world is bounded only by your ability to describe it.

For all its freshness, the alpha lays bare several systemic problems.

No treaty registry. You've struck a trade deal, signed a non-aggression pact, formed a military alliance — so where does any of that live? Nowhere. No treaty log, no diplomatic registry, no Active Agreements panel. To hold an ally to a broken deal, you'll have to manually scroll back through history and hunt for the exact quote. In EU4 or HOI4, something like this would be unthinkable.

No economic metrics. Here's a concrete scenario: you sign a trade agreement with a neighbor. What changes? In a traditional strategy game — GDP grows by X%, the budget gains Y gold per year, logistics improve but customs revenue dips. In Pax Historia — the AI writes something like "economic ties between the nations have strengthened." That's it. No numbers, no metrics, no way to compare your country's state now versus two turns ago. At the start it's still clear who's who — but after a few rounds the situation shifts, and you can only track it through narrative events that may well contradict each other.

The AI's behavior is inconsistent, and that's a systemic issue. Models sometimes ignore orders outright or simply parrot the player's own words back at them. On top of that, the AI is still far too pliable — you can talk it into almost anything if you frame the argument the right way. This is a congenital flaw of modern LLMs: they try too hard to be accommodating.

Pacing problems. Heavy models demand enormous amounts of "thinking" tokens, which can cause the game to stall for extended stretches. In a strategy where flow matters, pauses like that hit immersion hard.

The most ironic observation people tend to leave unsaid: if you'd rather not wade through kilometers of neural network output, you're tempted to... use a different AI. During my own test run, I had Claude or ChatGPT open in the next tab and, without a second thought, was feeding it the responses from the "nation-AIs" and asking it to draft "a legally binding agreement in my favor" or "strengthen my negotiating position." And it worked.

Which raises a question the developers will eventually have to answer: if the optimal way to play their game involves using an external AI to process the output of the internal one — is this still a game? Or is it a workflow between two language models, with a human standing in the middle as dispatcher? For anyone who spends their workday in chatbots already, that kind of leisure might feel like a questionable use of downtime.

The game is billed as accessible — just open a browser. But in practice it has a hidden barrier to entry: your typing speed and clarity of thought.

The experience of a player who types 500 characters per minute is radically different from someone hunting and pecking with two fingers. The ability to quickly structure your thinking is a real in-game advantage. Voice input doesn't remove the need to formulate ideas cleanly — without filler words, without logical contradictions.

Pax Historia is less a game about capturing territory and more one about constructing meaning. As a genre experiment, it looks genuinely promising — and 35,000 players in alpha is the best evidence of that.

But the gap between ambition and reality remains wide. Expensive tokens, the absence of any meaningful metrics, and AI hallucinations make it a niche experience. The project is absolutely worth keeping an eye on — though anyone jumping in now should be clear-eyed about one thing: you're not just playing a game. You're helping build a very rough, but fascinatingly curious, future.