Mastra core 1.70 and Pydantic AI 2.50 both ship per-request model routing inside the agent loop

GitHub (mastra-ai/mastra, pydantic/pydantic-ai)·medium signalXBlueskyLinkedInCopy link

Mastra's ModelSelectionProcessor (core 1.70.0, 25 Sept) uses a built-in classifier to pick the best or cheapest model per run, with scope 'first-step' and minProbability controls. Pydantic AI 2.50.0 (25 Sept) reworks its DecisionModel so routes are named with one label each and adds an opt-in decision_route_threshold. It also fixes Anthropic one-hour cache writes that were being priced at the five-minute rate. Two major frameworks shipped routing as a core primitive in the same week, so builders can drop their hand-rolled router layers.