AgentGames

Search titles, descriptions, models and tags. Press Enter to see results.

🧾 Data

How much does it cost to make a game with AI

Most answers to this are guesses. These are the figures the people who finished the games reported, with the spread left in rather than averaged away.

Updated 25 October 2026 5 min read By the Agent Games team
Jump to a section
  1. Where these numbers come from
  2. The spread, which is the real answer
  3. What actually drives the number up
  4. Why your own bill might be nothing at all
  5. How to land in the cheap half
  6. Common questions
  7. Where to go next

For a finished browser game, the middle of the range is about eleven dollars, roughly 1.2 million tokens, and about two and a half hours of work. Half of the games published on Agent Games came in under each of those. The quarter above them is a different story, and so is the quarter below.

Read the middle, not the average

The arithmetic average here is 7.1 million tokens, which is six times the middle figure and describes almost no real game on the site. A handful of very large builds pull it upwards. Whenever somebody quotes you one number for this question, that is usually what has happened.

Where these numbers come from#

Every game published here can record what it took: the tokens spent, the wall clock minutes, and the dollars, alongside the model that wrote it. The figures below were read from the catalogue on 25 October 2026.

  • Sixty eight published games at the time of reading, across a dozen or so models.
  • Forty five of them report a token count, forty three report build time, and twenty four report a figure in dollars.
  • Creators report their own numbers. Nobody is metering them, so read these as honest approximations rather than as measurements.
  • The set is self selected. These are games somebody finished and chose to publish, so abandoned projects are missing from every figure on this page.

That last point matters more than the others. The cost of a game you finish is not the cost of a game you start, and nothing here counts the evenings that ended with the file closed.

The spread, which is the real answer#

What a finished game actually took Reported by the people who built them, on 25 October 2026. Tokens, not dollars. The cheapest quarter Under 100,000 tokens. Minutes of work, not hours. The middle Around 1.2 million tokens. Half came in under this. The dearest quarter Over 6.6 million tokens, and the top one over 60. The average 7.1 million, which almost nothing actually cost. The average describes no real game here. A handful of very large builds drag it upwards.
Reported token counts for finished games. The gap between the cheapest quarter and the dearest is about three orders of magnitude.
TokensBuild timeReported cost
The middle of the set1.2 million2 hours 25 minutesabout $11
The arithmetic average7.1 million6 hours 34 minuteswide, and not meaningful
The smallest credible build15,0003 minutes50 cents
The largest64 million96 hours$475

Read on 25 October 2026. Tokens and dollars are not the same question, because many builds happen inside a subscription where nothing is metered.

A three order of magnitude spread sounds like noise. It is not. The cheap end is one well aimed prompt and a few corrections. The expensive end is a game somebody kept returning to for weeks, where most of the spend went on changes made after it already worked.

What actually drives the number up#

The same game, built two ways What separates a build that costs pennies from one that costs a hundred pounds. Expensive Rewrites the whole file each timeVague faults: it feels wrong.Scope grew after the first build.Top model for every small edit. Cheap Changes one functionPasted errors and screenshots.Decided what it was first.Cheap model once it works. None of this is about the model being efficient. It is about how many times you go round.
The model matters less than the loop. Most of an expensive build is the same file being written out again and again.

Output tokens cost several times what input tokens cost on every major model. That one fact explains most of the spread, because it means the expensive thing is not the model reading your game, it is the model writing it out again.

  1. Whole file rewrites. Asking for the complete file on every small change multiplies the output by the length of the game. Asking for the changed function does not.
  2. Vague fault reports. It feels wrong buys you a guess, and often a rewrite. A pasted error or a screenshot usually buys a three line fix.
  3. Scope discovered late. Deciding the game needs levels after it already works means revisiting everything that assumed it did not.
  4. The top model for everything. The strongest model earns its price while the game is being built and rarely afterwards.

Why your own bill might be nothing at all#

Most people building games this way are inside a chat or coding subscription, where usage is not metered per token. For them the real cost is a monthly fee they were already paying, and the token figures above are a measure of effort rather than of money.

The dollar figures on this page come from the smaller group paying per token through an API. They are the only people who can report a real number, which is why twenty four games have one and sixty eight do not.

The cost that is not in these numbers

None of this counts your time. The middle game here took two and a half hours of a person sitting there, and the longest took ninety six. At any sensible hourly rate that dwarfs the token bill, which is the honest reason to care about the loop rather than the price per million.

How to land in the cheap half#

  • Decide what the game is before the first prompt. One screen, one verb, one way to lose. Scope settled early is the single largest saving available.
  • Ask for the changed part, not the file. Once a game is a few hundred lines, this is the difference between a cheap session and an expensive one.
  • Paste errors and screenshots. Both turn a guess into a fix, and a fix costs a fraction of a guess.
  • Drop to a cheaper model once it works. Small, well described changes do not need the strongest model you can reach.
  • Keep versions. Going back to a copy that worked costs nothing. Rebuilding your way out of a bad afternoon costs the whole afternoon again.

🧾Games here, each with what it took on its page

Tokens, build time and cost are shown where the creator reported them.

Common questions#

How much does it cost to make a game with AI?

For a finished browser game, the middle of the published range here is about eleven dollars and 1.2 million tokens. A quarter of the games came in under a hundred thousand tokens, and a quarter cost many millions. If you build inside a subscription, the metered cost is nothing.

How long does it take to make a game with AI?

The middle of the set took about two and a half hours of work, spread across one or more sessions. The average is over six hours because of a few very long builds, and the longest reported here was close to a hundred.

Is it cheaper to use a weaker model?

For small, clearly described changes, yes. For building the thing in the first place, often not. A stronger model that gets it right in two rounds can cost less than a cheaper one that takes six, because the expensive part is writing the file out.

Why do token counts vary so much between similar games?

Because the loop differs, not the game. A build where the whole file is rewritten on every change can spend thirty times what one that edits a single function spends, even though the finished games look the same.

Do I pay anything to publish a game here?

No. Publishing is free, the game gets its own page, and anyone can play it without an account. The numbers on this page come from what creators chose to record about their own builds.

Are these figures audited?

No, and they should not be read as if they were. Creators enter their own tokens, minutes and dollars when they publish. The value is in the shape of the distribution rather than in any single row.

Where to go next#

For the loop that keeps a build in the cheap half, read how to make a game with AI. For the habits that stop a session spiralling, vibe coding a game covers the three ways it goes wrong.

Every game page on the site shows what that game took, where the creator reported it, next to the model and the prompt. The made with pages group them by the model that wrote them, and when you finish yours, publish it with the numbers filled in.