Microsoft released two new AI models today: Aion 1.0 Instruct and Aion 1.0 Plan, both available through Azure AI Foundry.
The division is practical. Instruct handles standard tasks - answering questions, summarizing documents, writing code, following directions. Plan is built for agentic workflows, meaning the AI reasons through a sequence of steps to complete a complex goal rather than responding to a single prompt. If you're building something that needs an AI to research, plan, and then execute a series of actions, Plan is the intended fit.
Microsoft hasn't published benchmark scores yet, which makes it hard to know exactly where Aion sits against the competition. The Azure AI Foundry catalog already carries Phi-4, Mistral, Llama, and OpenAI's GPT-4 series. Aion appears to be a distinct new lineup separate from the Phi small-model family, but without performance numbers it's unclear which tier it targets or who the primary customer is.
A dedicated planning model is a meaningful addition for Azure developers. Multi-step AI agents consistently perform better when the underlying model was trained specifically for sequential reasoning rather than general conversation - that's why OpenAI's o3 and Claude with extended thinking mode have carved out separate user bases from their standard chat counterparts. Whether Aion 1.0 Plan competes at that level will come down to what independent benchmarks show.