Azure AI Foundry provides access to various AI models. The right choice is crucial for success and costs.

Model Catalog

Azure OpenAI

GPT-4 Turbo (gpt-4-turbo)

• Strengths: Reasoning, complex tasks
• Context: 128K Tokens
• Costs: Medium-High
• Use Cases: Analysis, Coding, Consulting

GPT-4o

• Strengths: Multimodal, fast
• Context: 128K Tokens
• Costs: Medium
• Use Cases: Vision, Audio, Chat

GPT-4o-mini

• Strengths: Fast, cost-efficient
• Context: 128K Tokens
• Costs: Low
• Use Cases: Simple chats, classification

Meta Llama

Llama 3.1 (70B/405B)

• Strengths: Open-source, no vendor lock
• Context: 128K Tokens
• Costs: Variable
• Use Cases: On-premise, Custom fine-tuning

Llama 3.2 Vision

• Strengths: Image understanding
• Context: 128K Tokens
• Use Cases: Document analysis, image processing

Microsoft Phi

Phi-4

• Strengths: Small, efficient, reasoning
• Context: 16K Tokens
• Costs: Very low
• Use Cases: Edge, quick responses

Phi-3.5

• Strengths: Multilingual, vision
• Context: 128K Tokens
• Use Cases: Multilingual apps

Mistral

Mistral Large

• Strengths: European, GDPR compliant
• Context: 32K Tokens
• Costs: Medium
• Use Cases: EU compliance important

Decision Criteria

1. Task Complexity

• Simple (Classification): Phi-4, GPT-4o-mini
• Medium (Summarization): GPT-4o, Llama 3.1 70B
• Complex (Analysis, Code): GPT-4 Turbo, Llama 3.1 405B

2. Latency Requirements

• Real-time (<1s): Phi-4, GPT-4o-mini
• Normal (1-5s): GPT-4o, Llama 3.1 70B
• Tolerant (>5s): GPT-4 Turbo, Llama 405B

3. Costs

• Minimal: Phi-4
• Limited: GPT-4o-mini, Llama
• Flexible: GPT-4o, GPT-4 Turbo

4. Compliance

• Standard: All Azure models
• EU Focus: Mistral
• Full Control: Llama (Self-hosted)

Hybrid Strategies

Router Pattern

Requests are classified and routed to the appropriate model: Simple → Phi-4, Medium → GPT-4o, Complex → GPT-4 Turbo

Fallback Chain

In case of errors, it automatically switches to the next model: GPT-4o → Llama 3.1 → Phi-4

Cost Example

For 1 million tokens (Input + Output):

• Phi-4: approx. CHF 0.50
• GPT-4o-mini: approx. CHF 2.00
• GPT-4o: approx. CHF 15.00
• GPT-4 Turbo: approx. CHF 30.00

Prices as of January 2026, may vary

CNEXT Recommendation

Our experience shows:

1. 1Start small – Phi-4 or GPT-4o-mini
2. 2Measure – Track quality and costs
3. 3Scale smart – Larger models only where necessary
4. 4Test alternatives – Llama for specific cases

Conclusion

The choice of model affects quality, speed, and costs. CNEXT helps with the optimal strategy.