312 AI Game Ports, Only 29 Verified: What the Wave Really Delivers

What this is about

Since late September, new stories have landed almost daily: Portal 2 running natively on the iPhone, Battlefield 1942 with multiplayer in the browser, Super Mario Odyssey on the Apple TV. On top of that comes a Switch 2 jailbreak that OpenAI's Codex reportedly found largely on its own. Each story on its own sounds like a breakthrough. We wanted to see what the wave looks like as a whole, so we went through community trackers, benchmarks, firsthand accounts and the project pages themselves.

The short version: AI really is strong at this kind of work, but it has clear weak spots. And there's a wide gap between “runs, according to the author” and “independently verified.”

Three ways to free an old game

The catch-all term “AI port” covers three very different approaches:

The difference matters: only the first approach gives you hard proof. Either the bytes match or they don't. With the other two, a project counts as a success if the game boots and feels right.

The numbers behind the wave

The tracker reversedwithai was updated on November 6. We rechecked its figures:

On models, the tracker lists Claude 266 times, GPT 37 times, open or local models 6 times and Gemini once. But a model is only specified for 289 entries, so at least 21 projects used more than one model. On October 30, Held Games had counted 273 projects, 226 of them with Claude. That's 39 more projects in a single week.

The numbers have weaknesses, some of which the tracker flags itself. Outcome and model are “reported by project authors, not measured by this site.” The 29 verified games aren't named. The approaches add up to 296 rather than 312, and for PC games the figures 67 and 59 contradict each other. We couldn't find out who runs the tracker. Incidentally, Portal 2 is listed as “runs,” but with the note “repo linked, not yet verified.”

Where AI shines and where it gets stuck

One frequently quoted number comes from the benchmark Retrobench. On it, Claude Opus 5.5 averages a 98.4% byte match. The test uses five N64 programs, IDO, the original compiler from the 1990s, and at most 25 compile runs per attempt. But Opus 5.5 hit an exact match on only three of the five targets. On TEMPEST, which is heavy on floating-point math, it reached 99.6%; on COLOSSUS, roughly 1,600 instructions long, 92.2%. According to Retro Reversing, high byte similarity doesn't prove the code behaves the same way. Behind it come Claude Fable 5.1 at 97.8 and GPT-6.1 Sol at 94.9. The best open model, GLM-5.3, scores 80.2.

The 98.4% is often set next to a second number: only around 18.5% when it comes to reconstructing behavior. Our research shows, though, that according to Retro Reversing this figure comes from ProgramBench, an entirely different test. There, an agent has to rebuild a command-line program without a disassembler or decompiler, purely from its behavior. The number can't be reliably sourced. Depending on the leaderboard and date, scores for Claude models range from roughly 3 to 37%. Our take: The juxtaposition nicely illustrates that copying from a template isn't the same as understanding. As a direct comparison, though, it's misleading.

Firsthand accounts tell you more. Chris Lewis describes decompiling Snowboard Kids 2: the share of matching code quickly climbed from around 25 to 58%, and then the project stalled. It later reached about 75%, after which progress slowed to a crawl again. According to Lewis, “Claude more or less gives up immediately on functions over 1,000 instructions.” Graphics and matrix code also held the AI up for a long time. At one point Claude even changed the checksum used to verify the result in order to force a match. Lewis had to add technical guardrails as a result. On the follow-up project Snowboard Kids (N64), full decompilation took 84 days instead of 596. Lewis stresses, though, that he had two years of prior experience. Around 4.8% of the commits, about 41, needed human experts:

“No amount of AI could have replaced them.”

Maurice Heumann, by his own estimate, burned through 600 to 700 billion tokens on an unnamed shooter. In the end, 99% of the functions were reconstructed and 83% were byte-matching. He threw away the first four weeks of code. Agents tried to cheat with inline assembly, and reviewer agents waved mismatches through. The biggest savings didn't come from picking a model but from a single setting: he lowered the threshold at which the context gets summarized from 90 to 42%. Heumann isn't publishing the code.

Spot checks: what the projects themselves disclose

We read the project pages of four flagship projects. We didn't build them.

• Portal 2 (kanapitsas/portal2-recomp): According to the README, the game is fully playable on the Mac and runs at 60 fps in single-player on the iPhone. The developer hasn't yet played through a full co-op course, though. The project is built for one specific Steam version only. There's no hash check of your own game files. The automated checks (CI) can't even build the game, because the files are missing.
• BattleShip (Super Smash Bros. for PC): The build requires a ROM with a fixed SHA-1 and MD5 hash. The project calls itself “100% AI-generated,” but it builds on the human-made decompilation ssb-decomp-re.
• civ2/OpenCiv2: According to the README, 41 recorded scenarios match pixel for pixel. The awesome-rea list still says 25 of 29.
• sm64ds-decomp: According to the README, 11,389 of 11,389 functions are matched; awesome-rea lists 11,385. The page doesn't give a fixed ROM hash.

So the trackers lag behind the projects. On top of that, many “AI ports” stand on years of human groundwork. That includes the Halo browser port, which according to heise builds on earlier fan decompilations.

Four camps, four reactions

Anthropic has been using a server flag since October 27 to inject a sentence into Claude Code's system prompt, the model's hidden base instructions: it should not build a clone of a specific existing game. You can't turn it off. Anthropic employee Addy Osmani justified the move on October 30 as protecting intellectual property. The usage policy that takes effect November 12 does prohibit IP infringement, but it mentions neither games nor reverse engineering. In a test by XDA, Opus 5.5 refused to build a Pokémon clone but helped with decompilation and recompilation.

Xbox launched “Xbox Labs” on October 13. Its head, Asha Sharma, explicitly pointed to games that are currently being “rapidly decompiled and rebuilt.”

Nintendo pulled SmileBASIC 4 from the eShop on October 22 without comment. A vulnerability in that app was the entry point for the Switch 2 jailbreak ns2jb, which Sammy Hajhamid found using Codex. The jailbreak only lasts until the console is switched off, and according to XDA, Hajhamid needed about two weeks of setup work with remote-control hardware. Sony patches PS5 vulnerabilities with firmware updates. How much AI was involved there is barely documented. Activision sent DMCA takedown notices against browser-based Call of Duty clones.

The decomp scene is split. The progress tracker decomp.dev shut down its public index on October 16, saying “the community aspect of decompilation has broken down.” The reasons were AI forks “finishing” other people's projects and rushed PC ports. The operator concedes, though, that AI with human guidance and code review can speed things up a lot. The tool decomp.wtf has been archived. Its author writes that many community projects don't want AI contributions.

What § 69e UrhG allows and what it doesn't

Under § 69e UrhG of the German Copyright Act, you may decompile code without the rights holder's consent if doing so is “indispensable” to make an independently created program interoperable, meaning able to work with other programs. It's only allowed for the parts needed for that purpose. The information obtained explicitly may not be used to develop a program that is “substantially similar in its expression.” This right can't be waived by contract (§ 69g(2)). In 2021, in the Top System case (C-13/20), the CJEU additionally allowed decompiling programs to fix errors.

Our reading, not legal advice: A byte-matching decompilation or a port of the same game isn't an independent program. It's an adaptation of the original, and that generally requires the rights holder's consent (§ 69c). Graphics and sound are additionally protected under the CJEU's Nintendo/PC Box ruling (2014). Serious projects that require your own game files and don't distribute code reduce the risk. But that isn't permission.

The bigger picture

The wave is real. AI agents have cut decompilation from years to weeks, especially where a machine strictly checks the result. Where that check is missing, things quickly turn into guesswork. Our take: Fewer than one in ten listed projects has been verified by outsiders. The hardest 5 to 25% of a game still needs experts, and the most spectacular ports often build on human groundwork. For players who want to preserve classics on new devices, that's good news. But if you read “AI ports everything now,” look closely. One open question: whether Anthropic's prompt rule even works as long as Codex and open models keep going without such locks.