Traictory tracks 10 models from Microsoft. The earliest entry is Phi-3.5-vision-instruct from August 2024; the most recent is MAI-Thinking-1, released in June 2026. One of them shipped in the last twelve months.
The strongest Microsoft model in our data is Phi 4 Reasoning Plus, with an average benchmark score of 78.9% across 11 published results, including 68.9% on GPQA and 97.9% on FlenQA. Across the 10 Microsoft models with benchmark data the average is 69.0%.
3 of the models are available through an API, with input pricing from $0.05 to $0.10 per million tokens; Phi-4-multimodal-instruct is the cheapest. The largest context window is 128K tokens on Phi-4-multimodal-instruct. 2 of the 10 models accept images or other non-text input. 9 models ship with downloadable weights; the rest are API-only.
Microsoft Models
Other Companies
.svg?dpl=dpl_3EZdmE8NdzRwHesETrvSRCczw81T){kind=icon}
Microsoft model catalog updated in 2026. See model comparisons and best models for coding.