Claude Opus 5.5 Cuts a 1989 Game's Pixel Errors to Two

A developer running the same coding task past four successive AI models found Claude Opus 5.5 fixing a rendering bug down to two mismatched pixels, on a task earlier models never got running at all.

The developer, who writes as Priyan R, has spent since February feeding 1989's Apple II version of Prince of Persia to whichever frontier coding model was newest, asking each to port creator Jordan Mechner's published assembly source into a working C#/.NET port without the developer editing the code directly.

Where the earlier models failed

Claude Opus 4.6, tested in February, misidentified the game's frame-sequenced animation system as simple tile-hopping, according to the developer's write-up. The resulting build produced scrambled graphics and positioning errors that the model could not diagnose without being told directly what was wrong.

OpenAI's Codex, tested in March, fixed surface issues including pixel filtering, sprite transparency and gate collision detection. It did not question the flawed architecture underneath and never actually ran the game to check whether its fixes worked, per the write-up.

What changed with Opus 5

Claude Opus 5, tested in September, reverse-engineered the original DOS executable's Microsoft EXEPACK compression and located the game's animation tables through byte-pattern matching, rather than relying on guesswork. That let it extract real game data and produce the project's first genuinely playable build, identifying a ledge-detection bug by comparing its output against real screenshots of the original game.

Claude Opus 5.5, also tested in September, went further: it located and ported the room-drawing routine from SDLPoP, an existing open-source Prince of Persia reimplementation, and discovered that the game's brick pattern is seeded by room, row and column rather than randomized. In a single prompt, it cut a room-rendering discrepancy from 8,429 mismatched pixels to two, and caught and corrected its own sprite-positioning error after comparing frames.

Model	Tested	Result
Claude Opus 4.6	February 2026	Playable but architecturally wrong; misread animation system
OpenAI Codex	March 2026	Fixed surface bugs; never ran the game to verify
Claude Opus 5	September 2026	First truly playable build; reverse-engineered DOS format
Claude Opus 5.5	September 2026	Room-rendering error cut from 8,429 pixels to 2

The developer's own conclusion, laid out in the write-up, is that the biggest jump did not come from a smarter model alone. It came from giving Opus 5 and 5.5 tools to see their own output, through screenshots and automated comparisons against the real DOS binary, which let the models catch their own mistakes instead of needing a human to point them out.

What still doesn't work

The project's public repository, licensed GPL-3.0 because it incorporates code from SDLPoP, tracks every session in a file called SESSION_HISTORY.md, recording the exact model and prompts used at each stage. As of this month, the port has no guard AI, no sword combat, no sound and no palace levels built from the game's dungeon artwork.

The developer has not said which model will get the next attempt, or when. The project depends on the tester owning a legitimate copy of the original DOS game files for pixel comparison, which the repository does not distribute.

Anthropic cut Opus 5.5's own price alongside a benchmark comparison against its larger Fable 5.1 model earlier this month, while rival coding model Cognition's SWE-2 has been drawing its own price-versus-capability comparisons in the same market Opus 5.5 competes in.