Glm 5.1 Test : Making a Retro Style Game

GLM-5.1 builds prettier screens but MiniMax-M2.7 finishes the job

## The test setup The creator ran the same Westworld-inspired roguelike spec in parallel through Claude Code on two Chinese models: GLM-5.1 (z.ai, monthly light plan, $10, 90% weekly quota already used) and MiniMax-M2.7. Both settled on Phaser 3 + Vite + TypeScript, but their planning behavior diverged. GLM-5.1 "asked me the smarter questions" and went step-by-step through five phases; MiniMax-M2.7 "yolo'd it" and dispatched three agents in parallel. Web search was disabled on GLM-5.1 after the creator noticed it was "really slow" — even a by the way prompt stalled at 2.5k tokens for minutes.

## The speed gap After roughly 10 minutes MiniMax-M2.7 had already started dispatching agents and building a room generator while GLM-5.1 was "still on the scaffold project." By 25% of the 5-hour GLM quota, GLM-5.1 had consumed 6 million tokens and still hadn't delivered a playable build. MiniMax-M2.7 reached the debugging stage with one context compaction; GLM-5.1 had not yet compacted.

## The output quality GLM-5.1's loading screen and pixel-art tone were "impressive," but the build shipped with broken projectile physics, unhandled death, and misaligned UI. After two fix rounds, the creator said it "added glow to try to fix the projectile problem. But the projectiles don't actually move." MiniMax-M2.7 eventually rendered a brown room with heart icons and bullets that also didn't move, requiring 3–4 re-prompts before it loaded at all. Neither result was "acceptable."

## The Opus comparison An earlier Opus 4.5 run one-shot a working MMO with multiplayer, world chat, a minimap, and NPC proposal systems — but cost roughly $300–$600 on day one. The creator's call: GLM-5.1 is "nowhere near Opus level" for one-shot complex builds, and MiniMax-M2.7 still produces "a lot more errors" despite being faster.

## WildClaw benchmark Running GLM-5.1 on WildClaw's 60-test agentic suite scored 37% — above GLM-5 Turbo, possibly near MiniMax-M2.7 Pro — but "a lot of puzzle-solving tests" hit zero due to timeouts, not failures.