When the game was finished: how games were made in the 90s and what they compete for today

A game of the nineties had one property that today's games do not have: it was finished. Not because it was perfect, but because it ended up on a disc that went to the pressing plant. Once pressed and packed, nothing could be changed.

The original Dice Wars, the game this series comes from, belongs to the fading era of that world. It came out in 2006 as a Flash game by a single Japanese developer, and since August 2007 its Flash file has not changed, according to the web archive (more about that in the introductory article). That is a good reason to look at what that world looked like, and an even better reason to compare it with today.

Small teams and small budgets

The numbers from that time are surprising to today's eye. The table has only those that could be backed by a source:

Game	Year	What can be said
Super Mario Bros.	1985	32 KiB of program and 8 KiB of graphics, 40 KiB together; led by Miyamoto, about seven people
Civilization	1991	a cost estimate of 170,000 dollars (according to its author Sid Meier); Meier and Shelley
Doom	1993	the project began in November 1992 with five people; cost under a million dollars according to John Carmack's recollection of 2004
Diablo	1996	Blizzard North at the start about twelve to fifteen people; the authors hoped for approximately 100,000 copies, by November 1997 a million were sold
Half-Life	1998	an advance from Sierra of about a million dollars; the budget counted on about 180,000 copies
Operation Flashpoint	2001	Bohemia Interactive: the team grew from one to twelve permanent members and, according to the authors, was almost always understaffed
Mafia	2002	Illusion Softworks, Brno: four to five people at the start, 20 to 25 at the end (according to a developer in an interview of 2022; another source gives about 40)

Two remarks. The cost estimates are recollections of the authors, not accounting, and with Doom the recollection comes from a distance of eleven years. And the Slovak story is not missing: Wikipedia calls Spellcross (1997, Cauldron), according to Peter Poláček, the most significant game from Slovakia.

The data carriers were just as limited. The "1.44 MB" floppy disk actually holds 1,440 KiB, that is 1,474,560 bytes. A CD-ROM held 650 MiB. Everything the game needed had to fit on the carrier, and the carrier had to be made first.

Finished meant finished

What did a release look like? The last version was sent to the manufacturer (it was called the gold master), then it was pressed, packed and shipped, which took weeks. Patches in the history of software were sent out on disks and later on CD-ROM by post. Automatic updates of operating systems arrived only around the year 2000.

A note on accuracy: it was not entirely without fixes. As early as 1993 the magazine COMPUTE! wrote that the game Darklands by MicroProse had received six patches, after which a joint release "Darklands 7.0" came out. Those are exceptions, though; an ordinary game of that time did not change after release. The claim that patches commonly came on disks attached to magazines could not be confirmed: only recollections from discussion forums remained.

Testing was done by people. For Doom, before it was uploaded to the server, they tested for 30 hours in a row. For Diablo II the bug database held more than 8,300 records. The printed manual also served as copy protection: the game asked, for example, for a piece of information from a given page.

Several things from that time have also lasted for good. The famous Konami code (the NES version of Gradius, 1986) was not a bug but an aid: the developer Kazuhisa Hashimoto found the game too hard when testing it. The code stayed in the game after it was finished, and they left it there because removing it could have introduced new bugs. And bugs were, in Pac-Man, level 256 with its broken half of the screen (a number overflow) and, in Super Mario Bros., the world "-1". Today a patch would fix them and an anecdotal entry in the change log would remain. Back then they stayed in the game as folklore.

Flash: the era of one person and one month

The game this series is about was not born in a box. It was born on the web, in Flash, at a time when the world of web games was at its peak: Dice Wars appeared in June 2006 and Kongregate launched four months later.

Flash began as FutureSplash in May 1996 and Macromedia bought it in December 1996. Portals grew around it: Newgrounds (1995), Miniclip (2001), Armor Games (2004), Kongregate (2006). In August 2010 Newgrounds had more than 2.2 million users and over 180,000 games and movies, most of them made by one person.

The economy of that time can be quantified. According to the magazine Game Developer, FlashGameLicense (launched in April 2008) brokered more than 830 licensing deals by February 2009, worth about 956,000 dollars in total, an average of about a thousand dollars per game. A game that took a month of work received 5,000 dollars in advance, an eight-day one 1,500. Boxhead was made in six weeks, Dino Run in seven months. A game could be the work of one person and one month.

The end came in 2017, when Adobe announced that it would stop developing and distributing Flash Player. Support ended on 31 December 2020, and from January 2021 browsers blocked its content. According to Wired, the Flashpoint archive saved roughly 38,000 games in February 2020, and the emulator Ruffle helps to run them without Flash.

The world today: more games, less time

Let us look at the other side. The number of games released on Steam grows every year:

Year	New games on Steam
2019	8,097
2020	about 9,600 to 10,000 (the sources differ)
2023	14,533 (January 2024)
2024	19,003 (January 2025)
2025	about 20,000 (SteamDB 20,008, GameDiscoverCo 20,558)

The exact numbers differ between the sources, because Valve removes games after a while. For example, the year 2024 had a count of 19,003 in January 2025 and in December 2025 only 18,558. And almost half stay out of attention: in mid-December 2025, of the 19,008 games of the year, 9,269 (about 49%) were ones with at most 10 reviews.

Time belongs mostly to old games

Attention, however, is not distributed evenly. According to Newzoo, players on PC and consoles spent 61% of their time in 2023 on games older than six years (in 2021 45%). A second report shows 57% for 2024. The latest Newzoo report (2026) uses a new method and gives 53% and 54%. Matthew Ball, who also takes into account yearly series such as Call of Duty, gives around 71 to 72%. Every source gives a different number, but they all say the same thing: more than half of the time belongs to games older than six years.

Concentration is even stricter. According to Newzoo, in 2025 the twenty most played games held 58% of gaming time on PC (in 2022 68%) and 79 games made up 80% of the time on PC.

A new game therefore competes not only with other new games. Five titles (Fortnite, Roblox, League of Legends, Minecraft and GTA V) together held 27% of all gaming time in 2023.

Mobile: the first day decides

On mobile the numbers are sharper. According to the GameAnalytics report for 2024 (more than 11,600 games): of the players who install a game, in the best quarter of games about 27% return on the second day. On the seventh day the median return is about 3.4 to 3.9% of players, and in 75% of projects the retention after 28 days is below 3%. The median session length is 5 to 6 minutes.

The industry's answer is the daily login with a reward. A study of Korean mobile games from 2025 (a preprint, qualitative) found it in 38 of the 40 best games. What does not follow from that is important: evidence from a controlled experiment that a daily login raises retention was not found in the sources used. It is a widespread tool, not a verified method.

Eight seconds and the goldfish

In every talk about attention a claim appears that human attention has fallen to eight seconds, less than a goldfish has. It is not true. It comes from a Microsoft Canada report of 2015, which took it from the Statistic Brain site without a source. The BBC took it apart in 2017, and psychologists said that the "average attention span" makes no sense, because it depends on the task. In a survey by King's College London from 2021, every second Briton believed it.

The real finding is different and interesting: the researcher Gloria Mark of the University of California, Irvine, measures how long a person stays at one screen before switching. In a study of 2003 it was about 150 seconds, in 2016 to 2020 about 47 seconds. It is not an ability to concentrate, it is a behaviour: we switch often.

And what competes with a game for that time? Short video. According to an Ipsos survey of January 2026 (cited by Ball), more than half of adults in the USA watch short video at least a few times a week, more precisely 62%. Mobile games are played by 45% and PC or console games by 32%.

What this means for a game that is not meant to be finished

Fine, so what?

Audamir does not want to compete with Fortnite and cannot. It wants to be a game for fifteen minutes: open it in a browser, with no installation, and when it ends, click "Next game". Principalities^[1] last, depending on the number of players, from 8 to 20 rounds, and you get to the next game with one click, similar to Civilization with its slogan "one more turn", which generations of players know.

Two design decisions follow from this comparison. The first: no pressure. The game has no timers that would threaten the player with a loss, and it is not sold for real money. The second: honesty about the chances. In the article on boxes with visible chances (forthcoming article no. 12) you will see that with every reward, before the purchase and after it, its probabilities are shown.

It is not entirely clean. The Puzzle of the day has a series of days with a bonus, which is a mild form of the very thing that the design otherwise avoids. It is a compromise and belongs among the open questions.

The finished game no longer exists

Finally, the most interesting transformation. The original Dice Wars is finished. Audamir will never be finished: the plan of extensions has eighteen chapters, which are added according to what the players and the testing show. In the nineties the cost of a mistake was measured in boxes that had to be returned. Today it is measured in the hours it takes to fix the mistake. That is why it is good to have tests that say whether something broke after a change, and one of the later articles of this series is about that.

"Finished" in 2026 is a decision, not a state.

Sources and notes

For every claim a source and a degree of certainty are given. I pick the strongest ones. The cost estimates are recollections of the authors.

Doom, Quake, Diablo, Half-Life, Civilization, Super Mario Bros.: Development of Doom (with references to Kushner's book Masters of Doom), Diablo, Half-Life, Civilization, Super Mario Bros..
Operation Flashpoint: postmortem by the Španěl brothers (Game Developer, January 2002; a reprint). Mafia: interview with Roman Hladík for the 20th anniversary (another source gives about 40 people). Spellcross: Wikipedia.
Floppy disk and CD: Floppy disk, CD-ROM. Patches: Patch (computing); Darklands and six patches: COMPUTE! magazine, issue 157. Konami code: Wikipedia.
Flash: Adobe Flash, Adobe announcement of 25 Jul 2017, FlashGameLicense in Game Developer, 9 Feb 2009, Flashpoint: Wired, February 2020.
Steam: GamingOnLinux on the year 2024, GamingOnLinux of December 2025, Chris Zukowski on the year 2025.
Playing time: Newzoo via GameDiscoverCo (data for 2023), the Newzoo report 2026 (a mirror of the document) and Matthew Ball, State of Video Gaming in 2026. The share of old games differs by method (53 to 72%), which is why all of them are given in the text.
Mobile and the daily login: GameAnalytics benchmarks 2025 via GameDevReports, arXiv 2504.10714 (a preprint, Korean games only).
The myth of eight seconds: BBC, 2017 (the page could not be opened, it was read through secondary summaries), King's College London, 2022, Gloria Mark, University of California.
Not found: the budgets of Operation Flashpoint and Mafia, the number of people in the Half-Life team, Polish teams of the nineties, evidence of patches on disks attached to magazines, controlled evidence of the effect of a daily reward. SteamDB was unavailable for the research, so the numbers for Steam are from secondary reports.