Anthropic implements AMD GPU input processing in 3 days with Claude

Anthropic said it brought a Claude model's input-processing stage to usable performance on AMD's MI355X GPU in roughly three days using its Claude Mythos preview, disclosing the experiment during its IPO roadshow. The scope covers only input processing—the step where the model reads user prompts and begins computation—not full inference or answer generation. Claude reportedly ported or wrote nearly all the required GPU kernels with limited human intervention, a result that matters because it suggests AMD hardware can be made viable for frontier-model workloads with far less manual engineering. Whether the same gains hold for full inference and production-scale serving remains the key question to watch.