Giving frontier AI coding agents 24 hours to write a complete software GBA emulator
Mechanize, Inc. is trying various LLM models to see how well they can make a Game Boy Advance emulator in Webassembly given various time constraints. Above are the May 2026 numbers.
Each model gets a Docker container with the Rust + wasm32 toolchain, the ABI specification, a BIOS stub, dev ROMs, and an oracle CLI, which is a black-box wrapper around Mesen2 that the model can run on any ROM with any input sequence to see Mesen2’s behavior. The model cannot read Mesen2’s source and does not have access to the internet. We run each model for 24 hours, and take checkpoints of its progress every 15 minutes.
According to their charts, Claude Opus 4.6 is the leader, beating other commercial models and hands down better than local models.
Ed note: I wonder how Fable and Mythos will do…
See the results at gbaeval.com and all the nitty gritty on GitHub.
— Source: Adafruit Industries – Makers, hackers, artists, designers and engineers! — Read More
