Microsoft releases Phi 4 small language models for Azure and Windows, cuts token pricing for enterprise workloads

Microsoft on Tuesday launched Phi 4 Mini and Phi 4 Small, new compact language models available in the Azure AI model catalog and the Windows Copilot Runtime, aiming to reduce inference costs and latency for enterprise copilots. The company said the models ship immediately in selected Azure regions and support on device execution on recent Windows NPUs as well as serverless endpoints in Azure.

Microsoft is pushing price as the headline change. The company listed Azure serverless pricing at 0.15 dollars per million input tokens and 0.60 dollars per million output tokens for Phi 4 Mini, and 0.30 dollars per million input tokens and 1.20 dollars per million output tokens for Phi 4 Small, which it said is a 40 percent reduction compared with Phi 3 tiers. Per second vCore pricing for dedicated throughput on Azure AI inference endpoints is unchanged, according to Microsoft.

Pricing and performance

Microsoft published evaluation results that show Phi 4 Mini scoring 74 percent on MMLU and 65 percent on GSM8K with a 128 thousand token context window, and Phi 4 Small reaching 79 percent on MMLU and 74 percent on GSM8K. The company compared Phi 4 Mini to open models such as Google Gemma 2 9B and Meta Llama 3.1 8B, which reported MMLU scores in the high 60s to mid 70s in 2024, citing vendor papers from Google and Meta.

Microsoft said Phi 4 Mini delivers median first token latency of 35 milliseconds and about 15 tokens per second on Azure NC H100 v5 instances, with quantized builds running in 4 gigabytes of memory for Windows NPUs. The Windows Copilot Runtime packages the models with ONNX Runtime and DirectML so developers can ship the same prompts across cloud and client.

Enterprise integrations and governance

The models are integrated into Teams, Dynamics 365 Copilot, Power Platform Copilot Studio, and GitHub Copilot for classification, summarization, and agent orchestration tasks that do not require large scale reasoning. Microsoft said customers can isolate traffic with Private Link, use Azure AI Content Safety policies, and keep data within the EU Data Boundary. Microsoft lists HIPAA eligibility and FedRAMP High support through Azure regions that already carry those certifications.

The release advances Microsoft’s dual track model strategy that combines access to OpenAI models in Azure OpenAI Service with first party models in Azure AI. Microsoft introduced the Phi family as small language models for constrained hardware in 2023 and 2024, and maintains a multiyear investment in OpenAI that independent reporting has placed at about 13 billion dollars.

For the broader market, the push to smaller, cheaper models underscores enterprise demand to reduce per request costs and improve latency while keeping sensitive data local. It adds pressure on providers of compact models such as Google with Gemma, Meta with Llama 3 series, Mistral with 7B and Mixtral, and Nvidia with NIM microservices, and it raises the bar for production RAG and agent workloads that can move from general purpose large models to task specific small ones.