Why frontier models ship in waves

OPENAI SHIPPED GPT-5.5 on Thursday — internally codenamed "Spud," its first fully retrained base model in over a year. Sixteen days earlier, Anthropic had quietly released Claude Mythos Preview to a restricted list of Project Glasswing partners. Coverage has treated these as two separate launches from two competing labs; the more accurate reading is that they are a single event, timed by silicon.

Each model is a first of its kind for its lab. GPT-5.5 was pre-trained on what OpenAI and Nvidia describe as the first 100,000-GPU Blackwell cluster, a system the two companies jointly brought up and which OpenAI has now committed to scale to more than ten gigawatts and millions of GPUs over the coming years. Mythos comes out of Anthropic's Project Rainier, a Trainium cluster that has crossed the one-million-chip mark and which this week absorbed another five gigawatts of Amazon commitments, with a new Trainium generation slated to come online by year-end. The benchmark tables in each launch blog post are the visible artefact; the silicon underneath is the story.

Refresh rate

But the timing is not a coincidence, and it is not — as most coverage has implied — a sudden breakthrough in algorithms or reasoning. It is the chip cycle. Hopper, the Nvidia architecture that produced GPT-5, had been tapped out for months; Jakub Pachocki, OpenAI's chief scientist, conceded on Thursday's press call that the past two years of model progress had been "surprisingly slow." That is a technical euphemism. The labs had ridden Hopper as far as it would go, and then waited.