Microsoft AI progress is achieved through a disciplined loop called "hill climbing," involving continuous improvement via compute, data, and evaluation. This concept applies to improving deployable model packages in measured steps, focusing on quality, latency, and cost. The model router in Foundry Models brings this hill climbing approach to the model selection process, moving beyond manual or custom routing tools. This latest release expands the model router's capabilities by increasing the pool of supported models and its regional availability. New additions to the model pool include Anthropic Claude Opus 4.8 and the GPT-5.6 family, while older models are deprecated. The model router is now available in 28 global regions and 21 data zone regions to accommodate regulatory and governance requirements. Updates to the model pool occur automatically, maintaining a stable endpoint so applications do not need redeployment. The model router supports an optimization loop through A/B testing, model decomposition, and continuous routing. A/B testing allows teams to compare models and routing strategies for optimal production use. Model decomposition helps understand workload behavior and identify opportunities for specialization. Continuous routing enables per-request selection of the best-fit model, turning model selection into an ongoing optimization process.