LMRank Showcase: Same Prompt, Every Model

Benchmark tables tell you how a model does on a curated set of questions. They do not show you what it builds when you hand it a real brief. The LMRank Showcase fills that gap: every model gets the same detailed prompt, produces one self-contained HTML file, and we publish the result so you can open it and judge it yourself.

The point is comparison under identical conditions. Same brief, same constraints, no retries picked by hand. If one model ships a working game and another ships a pretty screen with dead buttons, you see it in the browser, not in a score.

The three briefs

Each brief is long and specific on purpose. Vague prompts reward models that fill space with decoration; precise prompts reward models that follow requirements.

• Cambrian Open World: an explorable 3D seafloor set 518 million years ago, with five recognizable organisms (Anomalocaris, Opabinia, Hallucigenia, Marrella, trilobites), three zones, a field guide, and a HUD with a working timer.
• Quantum-Powered AI Lab: a landing page for a fictional AI company that uses quantum computing as a supporting technology. It needs a working model selector, a coding-harness demo, and a two-qubit widget whose probabilities have to sum to one.
• Safari Reserve Manager: a 3D management game with three species, two building types, two seeded events, and a seven-day win condition that links money, habitat, and animal welfare.

What the briefs actually test

The briefs are built around invariants: behaviors that either work or don't. A few examples:

• Determinism: the 3D briefs fix a random seed of 518. Restarting must reproduce the same world. Models that ignore the seed fail visibly.
• Linked systems: in the safari game, placing a watering hole must cost money, add upkeep, reduce elephant thirst, and raise habitat health. A button that only changes a number on screen doesn't count.
• Behavior, not looks: in the Cambrian world, Anomalocaris must actually chase a trilobite at least once per run, and at least two species must react when the player gets close.
• Correctness: the qubit widget's displayed probabilities have to be mathematically consistent, and the page has to work with a keyboard and respect reduced-motion settings.

This makes the Showcase a test of long-form instruction following and code correctness in one file: roughly the job of a front-end engineer handed a detailed ticket. It does not measure reasoning depth, factual knowledge, or tool use; the leaderboard score and benchmark data on each model page cover those.

How to read the results

Every result links to a live preview and to the model page it belongs to. When you compare two outputs, check a few things in order:

• Does it load at all, and does it degrade gracefully without WebGL?
• Do the stated invariants hold when you actually play with it?
• Is anything missing from the brief, such as a species, a zone, a section, or a control?
• Only then: is it good-looking and pleasant to use?

The first results came from DeepSeek V4 Pro, DeepSeek V4 Flash, MiMo-V2.5, and MiMo-V2.5-Pro. More models are added as they become available through OpenRouter, which is how we run every brief so that each model gets the same API path.

Why no single score

We tried a weighted creative score (visual quality, interactivity, completeness, creativity). It collapsed very different failures into the same number: a stunning scene with broken controls and a plain but fully working game could land within a tenth of a point of each other. For a developer choosing a model, those are opposite outcomes. Showing the artifact and the requirements side by side is more honest than averaging them away.

Takeaway

Use the Showcase as a tiebreaker, not a ranking. If two models are close on the leaderboard and price, open their Showcase outputs for the brief closest to your work. The one that respects the invariants is usually the one that will respect your specs too.

Related: Best Coding Models · All Showcase results