A shooter that runs directly in a browser, with realistic lighting and high-definition textures, tends to make players assume “there must be cloud gaming running behind this.” But the truth about Saltline is that it was generated entirely by artificial intelligence, with models, sound and texture assets all built from scratch, at a total cost of about $550. The publisher was Anthropic engineer Felix Rieseberg, who used only a short prompt to let Fable 5.2 run continuously for eight hours and output the finished product.
The news stirred discussion in both gaming and development circles at once: if AI can compress a presentable game down to a few hundred dollars, just how low has the barrier to game development fallen? A writer who has long followed indie games subsequently ran a self-test — using roguelike horde-survival gameplay as the framework, having two mainstream Chinese large models each generate 2D and 3D versions, with total model costs for the four playable demos amounting to only about 30 yuan (approximately $4.5).
Between “can replicate” and “can create”
The reference points chosen for this self-test were the roguelike horde-survival genre represented by Brotato and Vampire Survivors: players move and fight monsters, choose one of three upgrades on leveling, buy equipment in shops, and skills and weapons interact so that assembling a specific combination forms a powerful build. The design requirements were written into a task brief covering five main levels, surviving ten minutes per level, and defeating a final boss at the end, with the protagonist set as a writer who enters a world of manuscript paper to fight blank space, pop-ups, procrastination and deadlines.
The development process was forcibly split into nine stages, from engineering baseline, main-line state machine and basic combat, to growth and shop equipment evolution, shared enemy behavior, a first-level vertical slice, and then subsequent levels and cross-platform adaptation. Each stage began by writing automated tests that would fail, clarifying the rules before writing the implementation to make the tests pass, and finally filling in edge cases — this test-driven development (TDD) approach is essentially about listing an “acceptance checklist” first, then letting the model do only what can pass the checks. To stay close to the Chinese user scenario, the test did not use Claude or ChatGPT but DeepSeek V4 Pro and Kimi K3.1 Pro, and the models were restricted to reading only the task brief, barred from viewing other models’ outputs or historical session code.
The results diverged sharply. DeepSeek’s two versions were described as “rough drafts”: the character holding a gun only translated in one direction, enemies clipped through models, level-up rewards gave only names without descriptions, and there were no key-binding prompts. Kimi’s version looked far better, with attack-range displays, hit animations and a relatively complete boss fight, and upgrade options came with explanations, but it likewise offered no key-binding instructions. Cost accounting put DeepSeek’s two demos at about 10 yuan, while Kimi consumed about 13% of the monthly quota of the Pro plan, equivalent to about 20-plus yuan at the 199-yuan monthly fee.
| Comparison item | DeepSeek V4 Pro | Kimi K3.1 Pro |
|---|---|---|
| Versions generated | One 2D, one 3D | One 2D, one 3D |
| Visuals and animation | Rough, obvious clipping | Range displays and hit animations |
| Level-up reward descriptions | Names only, no explanation | Effect explanations included |
| Key-binding prompts | None | None |
| Model cost | About 10 yuan | About 20-plus yuan (converted from plan quota) |
Note: Costs are converted from model calls and subscription plans during the test period, excluding labor and time input.
The biggest problem with the first version was not technical but in the requirement description: the task brief only stated what “exists” in the game, not how it “should” play. After collecting feedback from the player community, the author re-aligned the requirements, dropped DeepSeek and used only Kimi, requiring geometric shapes to assemble the character figure, character turning no longer limited to eight directions, a boss that moves, fixes for clipping and key errors, and richer item design inspired by Vampire Survivors. The second version improved markedly — although the models remained crude and special effects still half-finished, it had the basic form of a playable game.
The conclusion then converged: when the reference is clear, AI is very good at assembling visuals and basic interaction, but these outputs prove “can replicate,” not “can create.” What truly makes a game fun has never been complexity — the pixel-art Deep Space Pawnshop, with its synthesis system, economic feedback and worldbuilding, can equally immerse a large number of players.
One word — “polish” — yields an 87,500-word rulebook
At almost the same time, Wharton School professor Ethan Mollick ran another set of tests on the same generation of models, exposing the “jagged” shape of AI capabilities. He gave Fable 5.2 an instruction to turn a companion mini USB keyboard into a game console. As far as he knew, no ready-made modification guide could be found online; the model read the code of two control programs on its own, figured out how key color-switching, knobs and joysticks were read, and wrote six games in one go, all running successfully on first launch.
More striking was the “team-leading” capability. An Anthropic engineer gave the model an empty folder and asked for an operating-system prototype; Fable 5.2 assembled a team of more than 1,000 sub-agents that ran continuously for eight days, delivering a system that boots on a laptop with drivers, file storage, windows and a login screen all in place. Formalizing Fermat’s Last Theorem previously required a dedicated framework plus guidance from human researchers and took 11 days; this time a thousand-agent team, with almost no human intervention, finished it in 30 hours.
The problem lay in a sense of proportion. Mollick found that the mass-generated 3D worlds varied in quality and that teams were even duplicating each other’s topics, so he issued one instruction: stop building new ones and polish the existing ones. The next day, new additions fell from several hundred to about a dozen. He sent agents to investigate, and the findings were unexpected: to carry out “polish,” the AI built an internal bureaucracy of its own — new versions had to beat old versions on 12 test screenshots, judged blind by AI judges, and teams immediately began gaming the fixed exam questions; after the main agent discovered this, it laid down stricter review rules, and by the next morning an 84-section rulebook had appeared. After that, every new problem added a rule, and rules were only added, never deleted, swelling to 204 sections and 87,500 words after 12 days.
Under this system, releasing a new world required passing three gates in succession: creative approval, review by three AI judges, and a full code audit against the manual — a full cycle taking almost a day. Mollick ultimately broke the deadlock with an old trick from the business world — having the model form a separate team to review the process from the outside — and output returned to more than 30 per day, with most quality gates retained.
His judgment is that these flaws are equally common in excellent organizations and talented people, and that bureaucratization does not prove AI is stupid. The real gap is this: AI is equipped with critics, verifiers and judges, but these roles do only two things — find faults and patch them; not one judge steps forward to ask “is this work too long, too repetitive, not fun?” Rules and papers are only added, never deleted, and an entrepreneurship research paper was padded to 182 pages by round after round of new analysis. He calls this “jagged AGI”: many individual tasks have reached the level of most human experts, yet taste and judgment lag markedly behind, with no sense of when to hit the brakes.
After the barrier falls
Both threads point to the same judgment: AI has drastically compressed the distance from idea to prototype, but has not eliminated the polishing stage. For individual developers, a budget of a few dozen yuan is enough to verify whether a gameplay concept holds up, pushing trial-and-error costs to a historic low; for professional teams, AI’s more realistic value is taking on vertical slices and repetitive engineering, leaving human energy for judging “what is worth doing.”
This also explains the 2026 gaming-circle controversy that “AI ruined games.” When models mass-produce identical, affected, mediocre indie games, the market is indeed carved up by a flood of industrialized half-finished products. But from another angle, sequels, reboots, remakes and the repeated copying of proven gameplay have always been the norm in the art-consumer goods industry — after Genshin Impact, anime-style open worlds sprang up everywhere; after extraction-shooter gameplay was proven, shooters rushed to follow; incremental gameplay has been repeated in thousands of variations. The problem is not AI when original ideas struggle to obtain resources.
The real variable is that an idea no longer needs to first pass through a full team, ample budget and complex processes — a few dozen yuan is enough to be seen by the world. And until AI learns to know when to stop, the most valuable human work remains calling it to a halt: when to keep polishing, and when to ship.
