OpenAI, Mistral, and Nvidia all shipped their smallest models this week — and far from being a concession, these tiny models are the linchpin of the agent revolution. GPT-5.4 nano costs pennies per million tokens. Mistral Small 4 activates just 6B of its 119B parameters per query. Nvidia rallied eight AI labs to build open models that run on desktop hardware. The industry is collectively betting that the next phase of AI depends not on building bigger brains, but on making intelligence cheap and fast enough to run everywhere, all the time, which is exactly what the agent era requires.