This started as a straightforward comparison. Two AI game engines, Star and BitMagic, both running on Claude Opus 5, both given the exact same nine-step build plan for a dungeon crawler. Same prompts, same order, only fix prompts allowed to differ when one engine broke something the other didn't.

It did not stay straightforward, and it turned out better than I expected. One engine delivered a genuinely impressive result: a fully working, mobile-optimised, co-op multiplayer dungeon crawler with a persistent base, a shop, a signature elemental mechanic, and enough momentum that I'm still building on it now. The other ran out of budget before it produced a single stable build. Here's how it went, prompt by prompt, dollar by dollar.

The Build Plan

Before either engine touched a prompt, I laid out nine stages meant to build a dungeon crawler from nothing into a full game with a persistent base, a gear shop, a signature totem mechanic, base upgrades, a trophy room, and co-op multiplayer at the end. Both engines got identical wording at every stage. The only allowed divergence was bug fixes, since the two tools were never going to break in the same places.

Prompt One: First Blood

Star was up first with the core prompt: third-person dungeon crawler, melee combat, a couple of slime enemies, WASD movement, health bar UI. Cost: $2. What came back genuinely impressed me. A playable arena fight, torches, stone tiles, a wave counter, a working slime encounter, full animation, all of it working together straight out of the gate for the price of a coffee.

BitMagic was still processing that same prompt while Star had already moved on to prompt two. When BitMagic finally did return a result, it had burned through roughly 1,248 sparks (starting balance was somewhere around 2,600) and the output wasn't playable. The map had no boundary, so you could walk straight off the edge. The slime enemies were too small to hit reliably and killed the player before a hit landed. The game crashed on death.

Star: one prompt, one working game. BitMagic: one prompt, a long wait, and a broken build.

Star Hits a Snag, Then Fixes It

Prompt two asked both engines to turn the single room into a proper level, start room, a few combat rooms, corridors, an exit that only opens once enemies are cleared, plus a minimap. Star's attempt ($4.72) added the rooms and the exit worked, but the new room layout collided with the leftover geometry from prompt one. The result was a mashup of old and new maps full of invisible walls.

The next prompt ($3.06) fixed about half of those invisible walls, but introduced a new problem: instead of the intended room, corridor, room, corridor structure, the whole level generated as one continuous corridor. Less broken, but still broken.

The fix after that ($3.68 more, bringing the total repair bill for this detour to $6.30 alone) finally nailed it. Working level structure, correct minimap, animation locked in, hit detection solid, and it even added simple sound effects nobody asked for.

What stands out here isn't that Star hit a snag, it's how fast it recovered. Three prompts, a handful of dollars, and a genuinely tricky layout bug went from broken to buttery smooth, with bonus sound design thrown in for free. Running total after four prompts: $13.02, and a level structure that held up for the rest of the project.

Meanwhile, BitMagic

While Star was iterating through its room-layout fixes in the space of a single evening, BitMagic was still working on prompt two. Not stuck, exactly, just extremely slow. It eventually completed that prompt after burning through another chunk of sparks (down to 606 remaining from 1,352), and the result did crash on load. Reloading lost the build entirely and reverted back to the broken prompt-one version. The sparks spent on that attempt were just gone, with nothing to show for them.

At that point I switched strategy and tried a fix prompt targeting the known bugs (open map, slime hitboxes, death crash) instead of pushing forward. That got the map boundaries fixed and made the slimes bigger, but hit detection was still broken and sparks were down to 394.

One more attempt at prompt two burned the account down to -62 sparks, the prompt didn't even finish, and the result was functionally identical to what I'd already seen. That's where I called it. Over 2,600 sparks spent, and BitMagic never produced a build that both worked and persisted. I'm not going to pretend that's an unfair way to end the test. Both engines got the same prompts and the same amount of patience. One of them just never got off the ground.

Star Builds Out the Game

With BitMagic out of the running, the rest of this is Star's story, and it's a good one.

Prompt five ($3.68) added a difficulty ramp, tougher enemies the deeper you go, plus a treasure chest and gold drops at the end of each level. Then, instead of following the prompt list in order, I inserted a styling pass early to lock the visual direction before building more systems on top of it, which turned out to be a smart sequencing call and also taught me an expensive lesson (more on that in a second).

The Base hub landed next ($3.88), a proper home area with a rug, torches, banners, and a portal back into the dungeon. Progression carried over cleanly too, with a rune circle mechanic added along the way: stand on it for a couple of seconds mid-run and it banks your gold and items back to base. Die before banking, or before clearing every room in the run, and you lose everything unbanked. That's a proper risk-reward loop, and it wasn't even the headline system yet.

The Armory followed ($4.97 plus a $2.31 continuation, since Star hit a checkpoint and needed a "continue" to finish the job), adding a blacksmith forge and shop for weapon upgrades. I added an extra prompt here that wasn't on the original list: a fake AI NPC at the forge, purely so new players would have a clear visual point of contact instead of an invisible interact zone. That cost $3.09 and was worth every penny. The NPC actually works the forge bench and stops to acknowledge you when you approach, which nobody asked for but genuinely adds to the feeling of being in a real place.

Then came the Totem Shrine, Star's answer to the signature-mechanic prompt. Fire totems that burn enemies over time, frost totems that stack a slow effect, both only obtainable from dungeon chests rather than bought with gold. It worked perfectly first try ($6.33), complete with its own flavour text and a proper themed UI panel.

The Styling Lesson

Here's a useful tip if you're budgeting your own AI game dev project, and one Star itself pointed me toward. The Fable-style aesthetic pass, warm painterly lighting, richer colour palette, storybook fantasy feel, cost $9.24 across two prompts and delivered a solid warmer colour grade across the whole game. Good work, just not the full material and texture overhaul I had in mind, so I went looking for a cheaper route to get there.

That's when I discovered Star's real strength for this kind of thing: it's just as happy working with assets you bring yourself. I manually sourced and uploaded my own assets instead of prompting for them from scratch. Eleven new characters, three of them player-selectable, four as shop NPCs, the rest as enemies, cost $1 to integrate. A second batch of assets to properly change the game's look cost another $1. Total: $2 for a bigger, more substantial visual overhaul, and Star slotted every asset in cleanly.

The takeaway: Star isn't just a generator, it's a genuinely capable integration tool too. If you've got your own assets, or can source decent ones cheaply, Star will happily build around them for a fraction of the cost of generating everything itself.

The $12 Totem. One quirk worth flagging: generating a single decorative totem statue, one static 3D object, came out at $12.28, more than five other bug fixes and a full mobile optimisation pass combined. It's a reminder that pure asset generation costs can spike in a way that systems and logic prompts don't, so it's worth budgeting a bit of headroom for the occasional outlier. The statue itself looked great, and as you'll see below, it ended up earning its keep.

Mobile, Bug Fixes, and the Totem System's Redemption Arc

Star picked up mobile support along the way too: full touch controls, forced landscape orientation for a better mobile experience, for just $1.06. From that point the game was playable on both desktop and mobile, complete with swipe-to-look, tap-to-strike, and a proper action wheel.

A run of small, cheap bug fixes followed: a camera fix (57 cents), save bugs (a dollar and eight cents), a gold save error (27 cents), a floating statue (42 cents), and clearing out some broken assets (15 cents). This is the other half of the cost story worth noting: most fixes on Star, once the core system is right, are genuinely cheap.

The pricey totem statue got its redemption though. A fix and expansion of the totem system cost $5.98, followed by $4.57 to finish it off properly. What started as a $12 static object turned into a fully functional elemental totem system worth the investment.

The Finish Line: Co-op

The last item on the original nine-prompt plan was the highest-complexity one by far: 2 to 4 player co-op, scaling enemy count and loot drops per player, with a lobby screen in the Base. This is the kind of feature that can sink entire dev teams for months, and Star took it on directly.

• $5.52 for the first pass, needed a continuation to finish
• $11.72 to continue, still needed more work
• $2.49 for the finished version, first test: players could join the same room, but animations were out of sync and the host couldn't start the game
• $3.49 for a fix pass: co-op working, but with some lag, sync drift, and a join-in-progress UI bug
• $1.42 for final fixes, tested and working

Total for co-op: $24.64. And with that, the full nine-prompt build plan was complete. A working dungeon crawler with progression, a persistent base, a gear shop, a signature totem mechanic, mobile support, and working multiplayer.

What Started As A Benchmark Became A Real Game

Here's the part I didn't expect going in. Somewhere along the way this stopped being just a cost comparison exercise and turned into a project I actually wanted to keep building. After hitting "complete," I kept going: a full menu system ($1.53 plus 79 cents to finish), making some objects solid (33 cents), and a furniture edit mode letting players move and edit furniture in their home base, the foundation for a future buyable and placeable furniture system.

That furniture system took a bit more effort than expected. Understanding the existing code to implement it properly cost $1.55, an early attempt at the core prompt cost $1.06 (which turned out to still be groundwork rather than the real implementation), and the actual furniture editor and placement feature landed for $1.64. A camera bug that came with it cost $1.20 to fix, and a dedicated first-person camera mode for the furniture editor specifically added another 67 cents.

Why I Didn't Top Up BitMagic

Worth being upfront about something here. Going in, I'd budgeted to top up both engines if they needed it, and I did exactly that for Star more than once along the way. I didn't do the same for BitMagic, and that wasn't an oversight.

It came down to what BitMagic actually showed me in the time it had. Not a single failure to launch, but a pattern of them: horrendously slow generation times, a crash that wiped an entire completed prompt back to the broken v1 build, and the sparks spent getting there gone along with it. By that point I'd lost faith that more money would have changed the outcome. It felt less like "one more top-up gets us there" and more like watching the same result play out again at my own expense.

That matters for context too: this isn't my first time using either platform. BitMagic has actually done right by me before. I built a Candy Crush-style match-3 clone on it that turned out great, and that only took about 15,000 sparks. BitMagic clearly has real strengths, it just wasn't built for a task with this much moving complexity, layered systems, persistent progression, a shop, an elemental mechanic, multiplayer, all stacked on top of each other. I wouldn't rule it out for a future project. For something simple at its core, it's a solid choice.

But for anything complex, Star is where I'd point people without hesitation. It's worth looking at the caliber of games the Star community has produced, and a huge percentage of those creators aren't developers. No coding background, no formal game design knowledge, just people building genuinely impressive things because the platform was built to carry that weight.

Where The Game Stands Now

Since finishing the original nine-prompt plan, the game has kept growing. The Hold (the renamed Base) now has named NPCs manning their stations, Ysolde the Adept and Bronn the Blacksmith among them, and its own upgrade path through base tiers. The Crucible run has expanded to 8 full levels, with named encounters like the Level 4 "Dark Casters" wave fight. Weapons carry real identity now too, with elemental effects like frost built directly into higher-tier gear.

Beyond the original plan, The Hold has picked up a Ledger (for upgrading the base itself, higher tiers mean better loot on offer at the vendors), a Trophy Hall, a Co-op Gate for jumping into multiplayer runs, and a Decorate station for rearranging furniture, all systems that grew naturally out of what started as a nine-step test build.

If you want to see it for yourself, The Crucible is playable right here.

The Final Numbers

BitMagic prices its credits as "sparks" rather than showing a running dollar total the way Star does, which is part of why tracking the spend was a guessing game throughout this test. For reference, BitMagic sells sparks in packs: 2,000 for $5.99, 4,800 for $11.99, and 10,000 for $23.99. That works out to somewhere between $0.0024 and $0.003 per spark depending on the pack size. BitMagic burned through roughly 2,662 sparks over the course of this test (its starting balance plus the overdraft it ended on), which converts to somewhere in the region of $6.40 to $8.00 in real money.

That's worth sitting with for a second. BitMagic's failure didn't come from overspending, it spent a fraction of what Star did and still never landed a working, persistent build. Star costs more, but every single dollar of it bought something real and lasting.

Star earned this result. It hit bugs along the way, same as any dev process would, but it fixed them fast, cheaply, and permanently, and it kept building momentum the whole time. By the end it had produced a complete, polished, mobile-ready co-op game with a real progression loop and a signature mechanic, and it did it fast enough that I ended up wanting to keep building on it long after the comparison was technically over. That's a genuinely strong showing. BitMagic, by contrast, never got past the starting line, and the wait times alone would be a dealbreaker even before the crashes and lost progress come into it.

What has your experience been like with AI game engines? Have you had better luck with BitMagic, or hit your own expensive surprises on Star? Drop it in the comments, I'd genuinely like to compare notes.