The workhorse era: Hydrogen, decision models, and the cheap-LLM stack

In two days, the LLM market changed its shape. Google shipped Gemini 4 Hydrogen, a workhorse model that outscores its bigger sibling and costs $1.50 per million input tokens. The same week, Google and Perplexity both shipped models that do not write text at all — they decide. A 145-billion-parameter model learned to run on one laptop GPU, a 45-language model learned to run on a phone, and Anthropic let enterprises run the whole agent service inside their own walls. The frontier is no longer the story. The workhorse is.