Azure AI Foundry Reaches 1,600+ Models — Llama 3.3, Phi-4, and DeepSeek R1 Available as Pay-Per-Token Endpoints

What is Azure AI Foundry?

Azure AI Foundry (formerly Azure AI Studio) is Microsoft's centralised platform for deploying, evaluating, and managing AI models within the Azure ecosystem. It provides serverless API endpoints for LLMs — you call the model via a standard API without managing any GPU infrastructure. All requests stay within Azure's network, are covered by Azure's enterprise compliance certifications (SOC 2, ISO 27001, HIPAA, FedRAMP), and are controlled by Azure IAM policies.

What Changed?

Azure AI Foundry expanded from approximately 400 to over 1,600 models in its catalogue, including Llama 3.3 70B (Meta's latest instruction-tuned model at $0.00053/1K tokens), Phi-4 (Microsoft's small 14B parameter model at $0.000013/1K tokens), Mistral Large 2, DeepSeek R1 (reasoning model), Cohere Command R+, and Jamba 1.5. All are available as serverless pay-per-token endpoints — no GPU reservation or capacity commitment required.

Why Does This Matter?

Azure AI Foundry is now the most comprehensive managed AI model marketplace in cloud. The pricing is competitive: Llama 3.3 70B at $0.00053/1K is 26% cheaper than AWS Bedrock equivalent ($0.00072/1K) and significantly cheaper than Azure OpenAI GPT-4o ($0.0025/1K). The Phi-4 small model at $0.000013/1K is particularly remarkable — it handles classification, extraction, translation, and simple reasoning tasks at near-zero cost, making it viable for extremely high-volume automation pipelines that would be cost-prohibitive with larger models.

Model Selection Guide

Use Phi-4 ($0.000013/1K) for: email classification, sentiment analysis, data extraction from structured text, simple Q&A over short documents. Use Llama 3.3 70B ($0.00053/1K) for: complex document analysis, code generation, nuanced instruction following, multi-step reasoning. Use DeepSeek R1 for: mathematical reasoning, logical deduction, step-by-step problem solving. Reserve GPT-4o ($0.0025/1K) for: tasks requiring the absolute highest quality or native multimodal input (images). This tiering strategy reduces average AI costs by 60-70% vs defaulting everything to GPT-4o.

Who Should Act Now

Azure-native organisations should audit their current Azure OpenAI usage and identify which calls could route to cheaper models. A common finding: 60-70% of GPT-4o API calls are for tasks that Llama 3.3 or Phi-4 handles equally well at 5-20× lower cost. Implement model tiering with a simple routing layer: classify the complexity of each request, route simple tasks to Phi-4, moderate tasks to Llama 3.3, and reserve GPT-4o for only the most demanding use cases.