GPT-6 Astra cracks 99% of binary reverse-engineering tasks

OpenAI's GPT-6 Astra can read a compiled program and work out what it does without the source code. On SRE-Bench, a benchmark for reverse-engineering software binaries, it solved 88.0% of tasks on the first attempt and 99.2% within four. GPT-5.6 Sol managed 55.9% and 68.7%.

The viral claim is that AI now takes apart any game and rebuilds it. The mashup part is real: modders are merging gameplay systems from different titles, such as Minecraft mechanics in Elden Ring, with AI agents and decompilation tooling doing much of the grunt work. Open-source setups now hand Claude Code the whole loop of identifying the engine, decompiling, patching and testing.

Some of the posted claims are hype. Mario Kart 64 was fully decompiled by human volunteers in 2025, long before Astra. The post also says Claude Opus 5.5 does even better on SRE-Bench, and we found no published number for that. "Any software" is a stretch too. A benchmark score isn't a clean rebuild of Photoshop or an OS.

The analogy still holds. Source is the recipe and the binary is the cake, and models are getting good at working backwards from the cake. Decompiling games is mostly a hobby and a legal grey zone. For closed-source software vendors, that gap was a protection they leaned on for decades.

Why it matters: shipping a binary no longer hides how your software works, which hits anyone relying on closed-source secrecy or license enforcement.