AI Game Generators: The Mechanic Gap Nobody Prices In

An AI game generator will give you a playable browser game from one prompt. Whether it is worth paying for depends on a single thing: does the collision, scoring and state logic still work on the first play session? In a four-game benchmark with identical prompts — a multiplayer board game, a 2D platformer with double jump, a WebGL first-person shooter with WASD movement, and a Flappy Bird clone — exactly one of four platforms finished all four builds without a re-prompt. The other three lost a core mechanic: turn order, shooting, or score tracking. Art was never the failure point. Everything that broke was a rule.

Here is the disagreement that matters, and every figure below is a builder's own account rather than a measurement. One builder reports four builds finishing in minutes each, with no failures and no re-prompts, and says the winning platform was no slower than the rest. Other accounts count in months: one developer reports a shippable demo taking one to five months, with a staffed team of illustrators, modellers and a coder spending almost five months on a single title at roughly $1,200 a month in model access. A third developer gives two different figures for one build — roughly half a day in one breath, four hours in another — and reports about 50 tries on a single aerial attack. None of them is measuring the same job.

Money splits the same way, and the personal figures are not the same kind of claim as a published price. One developer reports that a $20 subscription produced seven mobile games, one of which he says took almost $400 on the App Store in two weeks and about $25 on Google Play. Another account puts a builder's running cost at around $400 a month; a third reports $600 spread over two months; and one of the accounts, conceding the builder never tracked the exact spend, puts the total at about $1,500 of subscriptions on a single game. Set those against the price lists. Lovable documents a free tier of 5 daily build credits, capped at 30 per calendar month, alongside a Pro tier at $25 and a Business tier at $50 at the 100-credit level. Base44 publishes five: Free at $0 with 25 message credits and 100 integration credits; Starter at $16 annually or $20 monthly for 100 messages and 2,000 integration credits; Builder at $40 or $50 for 250 and 10,000; Pro at $80 or $100 for 500 and 20,000; Elite at $160 or $200 for 1,200 and 50,000. Base44 states that annual billing saves about 20 percent. "A game" is not one unit of work.

What everyone agrees on is where the damage happens. Custom pixel-art assets get generated and then never wired into the game. A shooter loads its environment, accepts WASD, and has no gun. A platformer lets the player fall through platforms. A bird flaps, pipes spawn, and the score stops counting. In one entirely free 2D workflow the remedy is a fixed manual sequence: generate a character, hold that style constant across every later asset, crop and erase in a desktop photo app, upscale, isolate transparency with a Color Range selection, slice a sheet, offset and spot-heal the seams until textures tile, key a green screen out in Blender with a Keying node, export RGBA PNG frames, and begin with only a walk and an idle cycle because that is already enough to test gameplay. Green objects such as grass or trees need a blue backdrop instead, or the keyer eats them. Nobody generated their way out of that list.

The details worth stealing live in the workflows nobody else repeats. A sprite pipeline that registers a hit location per frame needs its own editor: mark a circle of influence on the fist, frame by frame, save the whole configuration as JSON, and let the coding agent read the file. That same build asked for a separate HTML visualiser purely to see which frames were playing, found one sheet running past the end of the animation, and then requested a tiny standalone positioning file rather than arguing with the agent about screen offsets. Set the offsets by hand. Finished. Generated motion drifts, too: a backward walk slowly pivots off centre, so the repair is a horizontal lock plus auto-centre before export. Code-synthesised sound failed outright after days of attempts; the workaround was to build a dedicated sound generator and request a handful of variations of every swing, hit and miss. Crowd movement is animated at a lower frame rate than the fighters on a wide backdrop. These are not model problems. They are pipeline problems, and the model will not build the pipeline unless you ask.

Consistency gets solved twice, in unrelated ways. One project writes every piece of project knowledge into a handoff document, closes the session, and opens a fresh one so the next agent inherits the same context. Another names each sprite sheet so the agent can infer state from the filename alone. Both are patches for the same weakness. A coding model sets up the state machine from that folder of named sheets; Nano Banana Pro holds the green backdrop steady across a wide character sheet; a separate procedural model makes worlds rather than meshes. Coding agents reach Blender through an MCP server, which Blender ships for 5.1 and later. Around them the industry is not neutral: 52 percent of developers say generative AI is hurting the game industry, and more than a third of the 7,000-plus Steam Next Fest demos disclosed AI use — 2,517 of them, 34.4 percent, against 750 and 21.2 percent in February. A roughly $400 million Series B at a $1.5 billion valuation sits behind one of the mesh generators. Tooling is industrialising faster than output is stabilising.

One last cost belongs on the invoice, not the feature list: building and running are metered separately. A Starter plan buys 100 messages and 2,000 integration credits, Builder buys 250 and 10,000, Pro buys 500 and 20,000, and a finished game left running keeps spending the second counter. Budget for the meter that keeps moving after you stop typing. Then ask for one mechanic that holds state across frames — a turn counter, a held block, a looping in-air kick — and play it again immediately. If the repeat run needs a re-prompt, the assets are irrelevant.