What it actually costs to build a game with an AI agent
Every public figure we could find, attributed — plus what the resulting games actually do when you run them.
The short answer
Nobody has published a clean number, and the honest range is wide: reported runs go from 40 minutes to a little over ten hours, and the only public token count is 1.3 million. What follows is every figure a named person actually reported, kept separate from our arithmetic and from our own measurements.
This question is asked constantly and answered nowhere: in a sweep of 100 of the 766 replies to the thread that started this (4.8M views), 11 asked what it cost — tokens, hours, dollars or rate limits — and the asking posts pulled up to 25,700 views each.
What builders reported
First-hand only — each person describing their own run. Self-reported and unverified; we did not measure any of this.
| @jaltucher | 10 hours · 1.3M tokens | Claude Opus 5 | Built the game from the published prompt |
|---|---|---|---|
| @runzhuotao | 2 hours | Claude Opus 5 | Same prompt, one-shot, on a laptop |
| @HelloVyom | 40 minutes · $30.62 | Fable | A DIFFERENT game (Jelly Jungle), comparing models — not the Claude of Duty prompt |
| @marckohlbrugge | $293.00 | — | A body of work over time, not one game |
| @mattshumer_ | no figure given | Claude Opus 5 | The original build; no figure given |
I applied Matt's prompt, ran it for a little over ten hours. About 1.3 million tokens on Opus 5. It created the game.
@jaltucher · 2,304 views
so i tried the prompt @mattshumer_ provided, and have opus 5 one shot the game for me, it took about 2 hours on my m5 macbook pro. and it's working!
@runzhuotao · 66,579 views
$30.62 dollars and 40 minutes for Fable to cook this!
@HelloVyom
pretty crazy what you can vibe code with gauntlet loop and just $293 in API credits
@marckohlbrugge
Still getting the final numbers but way less than you'd think. This is easily doable on a Max subscription.
@mattshumer_
What 1.3 million tokens costs
One report gives a token count. Converting it to money needs a number nobody stated — how many of those tokens were input and how many output — and on Opus 5 those are priced $5 and $25 per million. So the honest answer is an interval, not a figure:
$6.50 – $32.50 at list prices, 2026-08-08
All-input at the bottom, all-output at the top. A real agentic run sits low in that band and can fall below it: most tokens are context resent each turn, which prompt caching bills at a fraction of the input rate. This is our arithmetic on someone else’s reported token count — nobody published this dollar figure.
How far the guesses were off
Watching the thread, @_MaxBlade estimated 12 hours minimum. Someone who actually ran the prompt reported about two hours on a laptop. The spread between what onlookers assumed and what builders reported is most of why this question keeps getting asked.
We keep that estimate out of the table above on purpose: it is a bystander’s guess, not a run, and averaging it with first-hand reports would make the average worse, not better.
What you get for it
Cost is half a ratio. We loaded 48 of these games in a real browser and measured the other half — and it is the part the excitement leaves out.
| Median frame cost | 102 ms (~10 fps) | Time to draw one frame on a real GPU, sampled only after the game is open — not during its menu. A 60 Hz display allows 16.7 ms. Counts only the games we confirmed we got into. |
|---|---|---|
| Hold 30 fps | 7 of 24 | A game at 13 fps screenshots exactly like one at 60 — which is why the demos look better than the games. |
| Median download | 0.5 MB | Bytes fetched on first load. |
| Median load time | 2.3s | To the load event, desktop connection. |
| Fit a phone | 44 of 46 | No horizontal overflow at 390×844. |
Measured by us on 8 August 2026 — not reported by anyone. Every game, with its own numbers: the 56 AI-built games we catalogue.
Why the figures disagree
They are not measuring the same thing. One is a ten-hour agent loop on Opus 5; another is a two-hour one-shot of the same prompt; a third is a different game on a different model; a fourth is $293 of credits across a body of work, not one build. Scope, model, and harness each move the number more than the others, which is why a single headline figure would be worse than the range.
Prices used here: Claude Opus 5 at $5 input / $25 output per million tokens, as of 2026-08-08. Prices change; the arithmetic above is only as current as that date.
FAQ
How much does it cost to build a game with an AI coding agent?
There is no single price, and anyone quoting one is guessing. The public first-hand reports for the Claude of Duty prompt range from about two hours on a laptop to a little over ten hours and 1.3 million tokens on Opus 5. At list prices that token count works out between $6.50 and $32.50 depending on the input/output split, and prompt caching pushes a real run toward the bottom of that band. A separate report puts $30.62 and 40 minutes on a different game with a different model.
How many tokens does it take?
The one public token figure is about 1.3 million on Claude Opus 5 for a run of a little over ten hours, reported by @jaltucher. Nobody else in the thread published a count, which is exactly why the question keeps being asked.
How long does it take?
Reported: about 2 hours (@runzhuotao, on an M5 MacBook Pro), a little over 10 hours (@jaltucher), and 40 minutes for a smaller game on a different model (@HelloVyom). A bystander guessed 12 hours minimum — roughly six times the fastest reported run.
Can you do it on a Claude subscription instead of the API?
Matt Shumer, who built the original, said the cost was "way less than you'd think" and "easily doable on a Max subscription". That is his claim about his own run, not a measurement, and he did not publish the underlying numbers.
Are the games any good for the money?
That is the part almost nobody checks. We loaded 48 of them in a real browser on a real GPU and timed the 24 we could get past the menu into actual play: the median takes 102 ms to draw a frame, about 10 fps, and only 7 of 24 stay inside the 33 ms a 30 fps game is allowed. A game running at 13 fps screenshots identically to one running at 60, so demo videos systematically flatter this population. We got this wrong once already — an earlier version of these numbers was measured on a software renderer and partly on title screens, and we have published the correction.
Do these numbers apply to any game?
No. Almost every figure here comes from people running one specific published prompt for one browser FPS. A different genre, a different scope, or a different model moves the cost, and the reports themselves span three different models.
Method, and where this is weak
The cost figures are public statements by the people named, collected 2026-08-08 from one X thread and quoted with attribution. We did not verify any of them, and we have no access to anyone's billing. The dollar range is our arithmetic on a reported token count at published list prices — nobody published that figure. Only the performance numbers are ours: measured in a real headless browser on 8 August 2026, method and full dataset in how we measured these games. The sample is small, weighted toward one prompt, and skewed to whoever chose to post a number — treat it as the range of what has been reported, not as a market rate.