Anthropic reportedly used Claude Mythos Preview to bring prompt processing for one of its models to a useful performance level on AMD MI355X GPUs in about three days, according to Tom’s Hardware’s account of the investor-roadshow disclosure. Anthropic’s IPO page dates the start of its investor roadshow to November 10, 2026.
The result is about adapting software to run on a GPU platform. It does not measure how fast the MI355X runs inference or compare its speed with Nvidia GPUs.
The three-day figure is a software-onboarding timeline, not a measure of GPU performance.
What the three-day result covers
The reported task was prompt processing, also called prefill: handling the input tokens before a model generates its response. The account describes a useful performance level for this stage on MI355X GPUs. It does not describe a complete end-to-end workload; prefill and response generation, or decode, are distinct stages.
Anthropic also gave a historical estimate of six to 12 months for traditional engineering work to bring software onto new accelerators. That estimate comes from its past experience; it is not a controlled, like-for-like test against the reported three-day effort.
Why GPU software portability matters
A GPU kernel is a small program designed to run on a graphics processor. Adapting kernels is one part of getting software to work on a different accelerator platform. Nvidia’s CUDA is a programming platform; AMD’s ROCm software ecosystem includes HIP, and HIPify converts CUDA code for AMD’s platform, as Tom’s Hardware explains.
Anthropic’s reported timeline points to software onboarding as an area where AI assistance may reduce engineering effort. It says nothing by itself about whether the resulting workload runs faster on AMD or Nvidia hardware.