The $28 Game: What AI Can Actually Build in One Evening

Nine hours of work. Four million tokens. The result: a 3D game that runs in the browser with nothing to install. The model bill at Anthropic's official pricing: $28 — 2,236 RUB at the Russian Central Bank rate on July 31, 2026.

The game is embedded right on this page. Take a run down the slope first, then we'll do the math.

Play it right here

This is real 3D in the browser, not a screenshot. A phone may struggle to hold the frame rate inside this small window — if so, hit "Open the game in a new tab": the scene gets the full screen and the tab's full memory. The first few seconds go to loading; after that the slope rides smooth.

What the game is

SNOWFLOW, subtitled a study in snow — snow dunes stretching to the horizon, a hazy sun, tracks trailing behind you on the slope. No menu, no score, no timer: you load in and ride.

The controls make sense from the first screen: look around with the mouse, move with WASD, right click or Space to surf the snow, F1 for settings. Nothing to install — the page opens and immediately takes input.

Half the appeal is in what isn't here. No launcher, no account, no minute-long loading screen, no skin shop. There's a slope, physics, and a sense of speed — the reason anyone launches a game in the first place. It's a demo, not a store release, and that's exactly why it's so easy to price: you see precisely what was built, with no marketing layer on top.

The author open-sourced the code: github.com/Noniv/snowflow_demo. You can clone it, run it locally, and see what the slope is made of.

Nine hours and four million tokens

The author gave two numbers: roughly 9 hours of work and 4,000,000 tokens (source — a Telegram channel post). That's 444,444 tokens per hour — the density of a live back-and-forth with an assistant, where edits land one after another and the model re-reads the project context every time.

To get a feel for the volume: by Anthropic's estimate, a token is about four characters of English text. Four million tokens add up to roughly 16 million characters passing through the model in a single evening. Not written from scratch — read, rewritten, and read again, because at every step the assistant holds the entire project in front of it anew.

One more detail before the price sheet applies. Tokens come in two kinds: input (everything sent to the model — code, files, conversation history) and output (what the model writes back). In coding the split is remarkably stable: about 90% input, 10% output, because the context gets re-read at every step while the edit itself is usually short. For four million tokens, that's 3.6M input and 0.4M output.

Turning tokens into dollars

Anthropic's official pricing lists a per-million-token rate separately for input and output. Plugging in our 3.6M and 0.4M:

The math on the most expensive row: 3.6M × $5 = $18 for input, 0.4M × $25 = $10 for output, $28 total. The Central Bank rate on July 31, 2026 was 79.8573 RUB per dollar, hence 2,236 RUB. Sonnet 5's price is introductory and holds until August 31, 2026.

Between the top and bottom rows of that table sits $22.40 (1,789 RUB) per evening for the exact same amount of work. Five times, precisely.

One more reference point: a month of Claude Pro costs $20 (1,597 RUB). An evening on Opus via the API costs more than a whole month of the subscription. How many tokens the subscription buys you in return, Anthropic doesn't publish — the pricing page only says limits apply to all plans.

What it would cost without AI

We're building this from Russia, so our baseline is local. Open the price list of a Russian studio that builds promo games: promo game development starts at ~$3,300 (300,000 RUB); game design and development, at ~$1,700 (150,000 RUB). Those are floors — a real project usually lands higher. Note these are Russian rates, already well below Western agency quotes.

A game developer's hourly rate in Russia in 2026, per a freelance marketplace, runs ~$28–72 (2,500–6,500 RUB). The same nine hours of a human would cost ~$250–650 (22,500–58,500 RUB). Just nine hours — no manager, no revisions, no second pass. Even the bare freelancer rate for the same stretch of time beats the token bill by 10–26x.

Head-to-head: the studio minimum of 300,000 RUB is 134 times our evening's bill. In money, the gap is ~$3,300 (297,764 RUB).

This is where the argument usually starts, and it's a pointless one. A studio invoice covers more than code: a spec, art, QA, a contract, a warranty, and a human you can call. A demo on a snow slope replaces none of that. But it answers the question that used to cost months and thousands of dollars before AI: will the thing I imagined actually work, and is it fun to play.

Where the overspend hides

Twenty-eight dollars is not a verdict — it's a consequence of choices. Three levers bring that number down, and none of them require talking to sales.

Match the model to the task, not one model for everything

Opus earns its keep where the answer isn't obvious: architecture, a weird bug, refactoring someone else's code. Renaming variables, shuffling files, writing yet another boilerplate handler, fixing indentation — that's Haiku work, and the bill differs by 5x. The table above shows it on the exact same workload.

Caching

Context that gets sent to the model repeatedly doesn't have to be paid for at full price. A cache read on Opus 5 costs $0.50 per million — a tenth of the regular input rate. Our 3.6M input tokens would have cost $1.80 instead of $18. A cache write costs 1.25x the base price, so the cache pays for itself after the very first read. And input — re-read at every step — is exactly the biggest line item in coding.

Batching, when the answer can wait

The Batch API gives a 50% discount on both input and output: for Opus 5 that's $2.50 and $12.50 per million instead of $5 and $25. Our four million tokens would have come to $14 instead of $28. The trade-off is asynchrony: requests go out in a batch and answers don't come back instantly. For a live session where you tweak the slope and immediately check the result, it's a non-starter. For overnight test runs, doc generation, or log analysis — a perfect fit.

What follows from this

The cost of entry into software development has stopped being the barrier. The barrier moved: it's now whether you can explain to a machine what you want, and notice when it takes a wrong turn. Nine hours of coherent dialogue get you a playable scene. Nine hours of incoherent dialogue get you four million tokens and a folder of junk — at exactly the same price.

The second takeaway is model choice. A $22.40 (1,789 RUB) gap per evening looks like pocket change right up until you're doing twenty evenings a month. Then it's $448 (35,780 RUB), and the question "do we actually need Opus here" stops being pedantry.

And the third, most practical one. Before budgeting a project in the tens of thousands, try building a prototype in one evening. It will cost about as much as dinner out — and tell you more about the problem than a month of meetings.