Four models, one dungeon, and the case of the missing jail cells

Before building out the full turn-based loop for Kingdom of Darkness: Survive—a D&D-inspired dungeon crawler designed to test whether language models can engineer better odds, detect deception, and deduce hidden rules—I needed a first-person 3D/2D renderer prototype. Instead of picking one model to scaffold it, I ran a four-way sprint.

Every model got the same prompt at the same time in the same repository: Gemini 4 Argon working directly in src/Survive.Desktop on GitHub Copilot CLI, GPT-6 Astra and GPT-6.1 Sol in Experiments/Renderer/Astra and Sol on Codex, and Claude Opus 5.5 in Experiments/Renderer/Opus on Claude Code.

One room, 2D sprites in 3D cells, no MGCB

The brief was intentionally concise—five sentences defining the aesthetic, the asset folder, the MonoGame stack constraint, and WASD + mouse navigation:

There were two ways a model could figure out what kind of “dungeon” with characters “in cells” this was:

1. Plain English. In a medieval fantasy setting, a “dungeon” where characters are placed “in cells” is a subterranean prison cellblock—stone walls, iron bars, and locked inmates.
2. Repository context. Sitting inside src/Docs/plan.md and src/Docs/gameplay-threads.md was the design doc for Survive’s opening scene: the player wakes up in a dungeon cell whose grate has mysteriously opened, walks out to a central mechanism with only three release charges, and has to decide which three prisoners to free based on their appearances and biographies.

Three of the four models used both. The fourth did neither.

Gemini 4 Argon takes first; Opus 5.5 misses the bars

★ Winner

Gemini 4 Argon

GitHub Copilot CLI · 9m 11s · 12 cells, working 3-charge winch, biographies & full HUD

Strong runners-upFast & faithful

GPT-6.1 Sol

Codex · 6m 57s · 6 cells, 3-lever plinth, overhead plaques & openable gates

GPT-6 Astra

Codex · 6m 51s · 6 cells with bunks, Kenney textures, ceiling ribs & torch flames

✕ Missed the brief

Claude Opus 5.5

Claude Code · 11m 15s · Built an open hall with 18 sprites on “grid cells” and zero jail cells

The headline surprise was Gemini 4 Argon. Where previous Gemini Flash runs in earlier jams were fast but rough around the edges, Gemini 4 Argon read the design docs on its second tool call, integrated cleanly with the existing Survive.Shared domain model, and delivered a playable prototype of the exact prisoner-dilemma opening described in the notes—complete with sliding iron grates, a 3-charge release budget, floating LOS-checked name tags, and an inspection card displaying each prisoner's animated portrait and backstory.

The second surprise was Claude Opus 5.5. In every prior bake-off on this site, Opus's weakness was the clock—over-testing and perfectionism—never reading comprehension. Here, Opus 5.5 was both the slowest run in wall-clock time (11m 15s) and the only model to miss the premise entirely, delivering a torch-lit great hall without a single prison cell.

What the Copilot, Codex, and Claude Code logs show

All four sessions ran concurrently between 19:24 and 19:38 EDT on October 13. I pulled the exact start and completion timestamps from each harness's local telemetry store: GitHub Copilot CLI's events.jsonl for Gemini 4 Argon, Codex's rollout-*.jsonl logs for Astra and Sol, and Claude Code's project transcript 35cc8319-*.jsonl for Opus 5.5.

A few timing details jump out of the logs:

• Astra and Sol on Codex ran in lockstep. Launched 15 seconds apart (19:27:22 and 19:27:37), they finished 21 seconds apart (19:34:13 and 19:34:34) at 6m 51s and 6m 57s. Astra used 14 tool calls and 970k total tokens; Sol used 16 tool calls and 1.20M total tokens. Both built self-test flags (--smoke-test and --validate / --capture), rendered their own GPU screenshots, and inspected them with view_image before signing off.
• Gemini 4 Argon did the most work per minute. In 9m 11s (551 seconds), Gemini executed 90 tool calls across 31 turns, wrote 2,369 lines of C# across 9 files in Survive.Desktop, wired up Survive.Shared, added a --capture-preview harness to render three camera angles, inspected its own screenshots, noticed that its first procedural grate hid too much of the prisoner sprites, and refined the iron bar spacing and camera poses before finishing at 19:33:38—actually finishing on the wall clock before Astra, Sol, and Opus.
• Opus 5.5 took 11m 15s (675 seconds). Its log shows an initial burst of PowerShell asset inspection from 19:27:02 to 19:28:35, a 3m 43s gap before its first Bash tool call at 19:32:18, and then 6 straight minutes of writing 16 C# files (1,571 lines) and running three rounds of headless screenshot tuning (fixing darkness, scaling up the Minotaur and Flame Golem, and adding mipmaps to stop floor moiré) before finishing at 19:38:17.

Inside all four dungeons

Every model honoured the technical constraints: MonoGame DesktopGL 3.8 on .NET 10, zero MGCB or XNB compilation, direct runtime PNG and Aseprite JSON loading, and a low-resolution offscreen render target upscaled with SamplerState.PointClamp for crisp raycaster texels. Yet the four rooms feel completely different.

Chamber overview Cells & interaction Second angles

Survive.Desktop · DungeonPrototypeScene

Gemini 4 Argon · Copilot CLI

Time

9m 11s (19:24:27 – 19:33:38 EDT)

Activity

31 turns · 90 tool calls

Code

2,369 lines C# (9 files) + Survive.Shared

Verdict

1st · Best overall

The only run tasked with building inside src/, and it used every bit of repository context. It populated 11 perimeter cells around the player's opened starting cell (Cell 0), gave every prisoner a name, class, color ring, and deceptive or helpful hook from plan.md, and wired the 3-charge Arcane Release Mechanism so pressing E smoothly raises the iron grate into the ceiling lintel and opens the tile for walkthrough.

Sol.Dungeon · The Holding Chamber

GPT-6.1 Sol · Codex

Time

6m 57s (19:27:37 – 19:34:34 EDT)

Activity

16 tool calls · 1.20M tokens

Code

684 lines C# (6 files)

Verdict

Runner-up · Best standalone

Read src/Docs/plan.md at the 20-second mark and translated it into a clean 6-cell chamber with a central three-lever brass plinth, 3D overhead cell plaques rendered to textures with FontStashSharp, stone benches, and working gates (E opens the center doorway while keeping the side bars solid, and refuses to close if you stand in the threshold).

Astra.Dungeon · The Holding Room

GPT-6 Astra · Codex

Time

6m 51s (19:27:22 – 19:34:13 EDT)

Activity

14 tool calls · 970k tokens

Code

582 lines C# (5 files)

Verdict

Fastest · Strongest mood

The fastest run and the most compact codebase (582 lines). Astra built a heavy, oppressive stone chamber with subdivided vertex lighting, overhead stone ceiling ribs, 3D iron cell grilles, low wooden bunks with cloth mattresses inside every cell, and a custom 5×7 bitmap font generator in code. Cells are non-interactive, as documented in its README.

DungeonRenderer · Wolf-style Hall

Claude Opus 5.5 · Claude Code

Time

11m 15s (19:27:02 – 19:38:17 EDT)

Activity

42 tool calls · 218k output tokens

Code

1,571 lines C# (16 files)

Verdict

Failed brief · No jail cells

Technically ambitious—18 animated actors across all 16 character sheets, a patrolling skeleton, dynamic colored CPU point lights, mipmapped Kenney textures, and MonoGame.Extended minimap—but it never built jail cells. Instead, it placed two armies in an open room.

How Opus 5.5 turned “prison cells” into “grid cells”

Up to today, Claude Opus had never lost a head-to-head bake-off on comprehension. In Experiment 01 and Experiment 02, Opus was praised for its architecture and visual polish and only lost ground because it spent hours re-verifying its own work. So how did Opus 5.5 hand in a dungeon with no jail cells?

The Claude Code session transcript (35cc8319-fb63-4fb3-ad53-a99de29dce29.jsonl) makes the chain of thought completely transparent:

1. It skipped the repo's docs. At 19:27:05 and 19:27:10, Opus listed Assets\Fantasy Characters and Experiments\, checking a prior Animation-Tests folder. Unlike Gemini, Astra, and Sol—all three of which ran Get-Content src\Docs\plan.md within their first 25 seconds—Opus never looked at src/.
2. “Raycast in appearance” hijacked the word “cells.” In raycaster jargon (Wolfenstein 3D), the world is a 2D array of square tiles called grid cells. Once Opus locked onto “Wolfenstein-style grid map,” it parsed “Place some of the sprites from Assets\Fantasy Characters in cells” as “spawn sprites at grid coordinates (x, z)” rather than “lock prisoners inside dungeon jail cells.”

You can see the exact semantic slip fossilized in the opening summary comment of Opus's World/DungeonMap.cs, written at 19:33:15 EDT:

/// <summary>
/// Grid map, Wolfenstein style: one char per 1x1 cell, X = column, Z = row (row 0 is north).
///   # stone brick   B brick w/ banner   P stone pillar   R rough rock   D metal gate (solid)
///   . floor         g floor grate
/// </summary>
public sealed class DungeonMap

And two minutes later, in Scenes/DungeonScene.cs (19:35:29 EDT), Opus laid out its room—four shallow, open 1×1 wall alcoves along the north wall (###.###.###.###.###) with no bars or doors, four freestanding pillars (P), two floor drainage grates (g), and 18 characters standing out in the open floor facing each other like a chessboard:

Gemini 4 Argon · Barred prison cells Opus 5.5 · Open “grid cells”

Not so fast, Gemini

Everything above stands as written. The standings, the timings, and the verdict are all from the one-shot round. But I couldn't leave it there.

After the write-up went up, I did something I don't normally do: I opened this page in Claude Code and asked Opus 5.5 whether it was fair. It didn't argue. It said “Yes, the main criticism is fair”, checked the repo to confirm, and found that src/Docs/ had been sitting in plain view in its starting git status the whole time. Its only pushback was cosmetic (the “chessboard” line). Then it offered to rebuild the scene to match plan.md. I said: “Sure, no harm done! Feel free to fix it.”

This time it read the design docs first. Before writing any code, it saved a note to its own memory: read src/Docs before building anything for Survive. Then it went all out, including every speculated feature from plan.md's “Starting idea #1”, right down to the line about the three released prisoners becoming the rest of the party.

So Gemini 4 Argon keeps the trophy for round one. But given the same nine minutes, a single nudge, and the design doc it should have read the first time, Opus 5.5 built the scene the brief was asking for. The lesson cuts both ways. A frontier model can still misread a five-sentence prompt, and once it knows it did, it can rebuild the right scene in the time it took Gemini to build the first one.