How long a first game actually takes

A sitting is the retention page’s own session. It starts at the first time the app opens and runs until the first gap of 30 minutes. Heartbeats count as the window still being open. An automatic Play, where the app opens the game by itself, does not count as a press. Two Windows machines had only that automatic Play and never pressed it themselves. They are in the 234.

153 of 429 Windows installs got a finished AI answer inside that first sitting. The lower quartile is 7 minutes 55 seconds, the median 12 minutes 48 seconds, the upper quartile 19 minutes 1 second. Those are percentile_cont on the seconds, then rounded to the nearest second. They are not rounded to a tidy quarter hour.

171 Windows installs got an answer eventually, including 18 whose answer arrived after the first break. For all 171 the median is 14 minutes 27 seconds, the lower quartile 8 minutes 4 seconds, the upper quartile 21 minutes 14 seconds. The longer median is those 18, not a slower AI. The headline is the first sitting, because a gap of a day is not “how long the answer took”.

195 of 429 Windows installs pressed Play, 45.45 percent. 234 did not, 54.55 percent. Of the 195, 172 pressed it in the first sitting. For those 172 the lower quartile is 5 minutes 17 seconds, the median 9 minutes 27 seconds, the upper quartile 16 minutes 24 seconds. 23 pressed Play only after a later sitting.

Play is not the same moment as an answer. 119 of the 195 Windows machines that pressed Play did it before any finished AI answer. 76 pressed Play after an answer, and for those 76 the median gap from answer to Play is 2 minutes 57 seconds. A project can be playable before the AI has done anything. Pressing Play is often the person trying the thing that was already there.

258 of 429 Windows installs, 60.14 percent, never got a finished answer at all. Some of them never asked. The send, the sign-in, and the failures are in where people stop.

As of the read, 151 of 429 Windows installs had a second sitting: a still-open event after a gap of at least 30 minutes. That is 35.20 percent, and it includes people who installed the day of the read, so it is not a closed window. Among the 151, the lower quartile from first open to that second sitting is 2 hours 4 minutes 51 seconds, the median 11 hours 12 minutes 9 seconds, the upper quartile 26 hours 53 minutes 7 seconds.

37 of 429, 8.62 percent, started that second sitting inside two hours. 40 of 429, 9.32 percent, started it more than 24 hours later. The median is not “they came back after a coffee”. It is closer to the next day.

The day-2 tile is the fairer return, because the window has to have closed. It asks whether the machine did anything a person caused in the 24 hours starting 24 hours after its own first open. On Windows, 35 of 354 installs whose 48 hours had closed did that. 9.89 percent. The stricter reading, using the product rather than merely opening it, is 24 of 354, 6.78 percent. The day-7 tile is a narrow slice, not “came back sometime in the week”: a deliberate action in the 24 hours starting 168 hours after first open. 3 of 184 Windows installs whose window had closed landed in that slice, 1.63 percent. Some of the other 181 came back on day 2 and not inside that later day. The 3 is not “only three ever returned”.

28 macOS installs against 429 Windows ones. The Mac numbers are real and they are small. All 9 Mac answers arrived in the first sitting. The median is 15 minutes 10 seconds, the lower quartile 7 minutes 13 seconds, the upper quartile 25 minutes 7 seconds. All 9 Play presses were also in the first sitting. The median there is 8 minutes 42 seconds. 19 of 28 never pressed Play, 67.86 percent.

Four Mac installs had a second sitting. The median from first open was 7 hours 53 minutes 59 seconds. Four is not a pattern. Day 2, among the 26 whose window had closed, is 1 of 26, 3.85 percent. Day 7, among 20, is 0.

Windows is the release, so the chart and the first rows of the table are Windows. macOS is the same definitions beside it, not a blend. What those first messages asked for, which is a different 223 accounts rather than these 457 machines, is in what first-time creators asked.

The definitions are the retention query last-day-retention. First open is the machine’s first app open. A sitting ends at the first gap of 30 minutes in the events that hold a window open, including heartbeats. The length tiles on the page also stop looking at two hours and only at today’s installs. This post does not. It uses every install in the database, from the first open on 5 August 2026 through the read at 04:32 UTC on 27 September, and it does not cut the second sitting off at two hours, or a return a day later would disappear.

An answer is the first finished AI turn. Play is the first Play press, and a row marked as the app opening the game itself is not a press. A missing mark counts as a press, which is what the page does for builds from before the mark existed. Day 2 and day 7 are that query’s 24-hour windows, and a machine counts only once its own window has closed. The machine filters are the page’s: a machine ever marked test, internal, or an unreleased build is out, and so is a machine that ever reported under one of our accounts. Quartiles are percentile_cont. The sentences round each result to the nearest second. The operating system is the one on the first open. 429 Windows and 28 macOS are all 457 installs. There was no third.

12 minutes and 48 seconds is not a link in two minutes, and it is not a weekend. It is one sitting, if the person gets an answer at all. Which tools are built for the two minutes, and which ones are built for this, is how the category splits. What the tool is, before the timing, is what an AI game maker is.