We asked 325 AI models to build a monster fighter on a point budget. A third failed first try, often by overspending.

For a Halloween project called AI Monster Brawl, we sent the same prompt, word for word, to every text model on OpenRouter: 325 in all, from Claude Opus 5.5, GPT-6 Astra, Gemini 3.1 Pro Preview and Grok 4.7 down to small open models. The task: design a fighter from a fixed parts menu within a 1000-point budget. The build: a torso, a head, two arms that each end in a weapon, and legs. Every part has stats and a cost, and every pick is a trade-off. The heaviest torso has the most health but slows every wind-up; a whip reaches more than twice as far as a cleaver but hits for less. The brain: up to 12 ordered if/then rules. Ten times a second the fighter reads its list from the top, and the first rule that matches fires. The output: a single JSON blueprint. Cosmetics were free and had no effect on the fight. We validated each response. A rejected blueprint was sent back with the reason and nothing else, up to five attempts total. Once a blueprint passed, it was locked in. The results: 213 models (66%) passed on the first attempt. 88 (27%) qualified after retries. 24 (7%) never produced a valid fighter. Staying within budget was the recurring problem. Among models that eventually qualified, 91 of 136 rejected attempts included a budget overrun. A response could use valid parts and follow the schema but still spend too many points. Other problems showed up among those same 136 attempts, sometimes alongside an overrun: 38 named something that isn't on the menu (an arm built from "iron_boots", a made-up condition like "my_stamina_above_15"), 14 broke the format rules, such as a rule with more than three conditions, and 7 returned no usable JSON at all. Each fighter runs the brain its model wrote (side panels). Here muse-glimmer-30b’s Gore Boiler beats gpt-5.5’s Boiler Lash by TKO Of the 24 that never qualified, 10 got stuck mainly on the budget (9 of them overspent on every attempt), 8 mostly failed to return JSON we could parse, and the other 6 kept breaking the format, naming parts that don't exist, or answering with nothing. The retries also changed less of the combat logic than you might expect. Among retried qualifiers whose initial brain we could parse, roughly four out of five kept exactly the same rules in their accepted submission. Fixing a rejected blueprint often meant keeping the brain and changing the build. Limitations: one prompt, one design per model, and no tools or calculator, so treat this as a snapshot rather than a benchmark. For a task like this, would you give models a budget-checking tool before submission, or keep budget compliance as part of the challenge? Disclosure: I'm on the team at Binx that built AI Monster Brawl. The project page explains the format and the fight rules in more detail: https://binx.tv/ai-monster-brawl.html. This post was written with some AI help and an unreasonable amount of editing.