Back to Blog

Azure AI Foundry Pricing and Cost Management Tips

April 9, 202612 min readMichael Ridland

Azure AI Foundry pricing is consumption-based, which sounds straightforward until you get your first invoice and wonder where the money went. We've helped dozens of Australian organisations plan and manage their Azure AI spend, and the gap between expected and actual costs is one of the most common problems we solve.

This guide breaks down exactly what costs what, shares real numbers from production workloads, and gives you practical strategies for keeping spend under control.

How Azure AI Foundry Pricing Works

There's no flat licence fee for Azure AI Foundry itself. You pay for the underlying Azure resources you consume. The main cost components are:

  1. Model inference (API calls to deployed models)
  2. Compute (for fine-tuning and managed deployments)
  3. Storage (training data, model artifacts, documents)
  4. Azure AI Search (if using RAG)
  5. Networking (data transfer, private endpoints)

Let's break each one down with current pricing in both USD and AUD.

Model Inference Costs - The Biggest Line Item

For most organisations, inference costs make up 50-80% of total Azure AI Foundry spend. This is the cost of calling your deployed models.

Pay-Per-Token Pricing (Serverless API)

Pricing is per million tokens. A token is roughly 4 characters or 0.75 words in English.

Model Input (per 1M tokens) Output (per 1M tokens) Notes
GPT-4o ~$2.50 USD / $3.80 AUD ~$10.00 USD / $15.20 AUD Best reasoning, highest cost
GPT-4o mini ~$0.15 USD / $0.23 AUD ~$0.60 USD / $0.91 AUD Best value for most tasks
o1 ~$15.00 USD / $22.80 AUD ~$60.00 USD / $91.20 AUD Advanced reasoning, very expensive
o3-mini ~$1.10 USD / $1.67 AUD ~$4.40 USD / $6.69 AUD Good reasoning at moderate cost
Llama 3.1 70B ~$0.27 USD / $0.41 AUD ~$0.27 USD / $0.41 AUD Open source, self-hosted option
Mistral Large ~$2.00 USD / $3.04 AUD ~$6.00 USD / $9.12 AUD Strong alternative to GPT-4o
Phi-3 Mini ~$0.13 USD / $0.20 AUD ~$0.13 USD / $0.20 AUD Smallest cost, simpler tasks

Prices are approximate and fluctuate. AUD conversion at 1.52. Check Azure pricing for current rates.

What these numbers mean in practice:

A customer service chatbot handling 1,000 conversations per day, with an average of 500 input tokens and 300 output tokens per conversation:

  • Using GPT-4o: ~$4.25 USD/day = ~$130 USD/month = ~$197 AUD/month
  • Using GPT-4o mini: ~$0.26 USD/day = ~$7.80 USD/month = ~$12 AUD/month

The difference between model choices is massive. We've seen clients reduce costs by 80% by switching appropriate workloads from GPT-4o to GPT-4o mini with no measurable quality loss.

Provisioned Throughput Pricing

For high-volume production workloads, Azure offers Provisioned Throughput Units (PTUs) - reserved capacity at a lower per-token rate.

  • You commit to a minimum capacity (measured in PTUs)
  • Pricing is per PTU per hour, not per token
  • Makes sense when you have consistent, predictable volume

When PTUs make financial sense: Generally when you're spending more than $5,000-$10,000 AUD/month on a single model's inference costs and your traffic is relatively steady. Below that threshold, pay-per-token is more cost-effective.

Compute Costs

Compute charges apply when you're fine-tuning models or running managed compute deployments (as opposed to serverless API).

Fine-Tuning Compute

VM Size Price per Hour (AUD approx) Typical Use
Standard_NC6s_v3 (1x V100) ~$4.50 Small fine-tuning jobs
Standard_NC12s_v3 (2x V100) ~$9.00 Medium fine-tuning jobs
Standard_NC24s_v3 (4x V100) ~$18.00 Large fine-tuning jobs
Standard_NC24ads_A100_v4 (1x A100) ~$5.50 Fast fine-tuning
Standard_ND96asr_v4 (8x A100) ~$44.00 Large model fine-tuning

Practical example: Fine-tuning GPT-4o mini on 500 training examples typically takes 20-40 minutes and costs $2-$5 AUD in compute. Fine-tuning GPT-4o on 1,000 examples takes 1-3 hours and costs $10-$30 AUD. These are low compared to inference costs over time.

Managed Compute for Inference

If you deploy open-source models (Llama, Mistral, Phi) on managed compute rather than through serverless API, you pay for the VM by the hour:

  • A single A100 GPU for serving Llama 3.1 70B: ~$5.50 AUD/hour = ~$4,000 AUD/month
  • A single V100 for serving smaller models (Phi-3, Llama 3.1 8B): ~$4.50 AUD/hour = ~$3,240 AUD/month

When managed compute makes sense: When you have very high volume (millions of tokens per day) and a single model to serve. The break-even point versus serverless API depends on your specific volume, but typically it's above 100 million tokens per month.

Azure AI Search Costs

If you're building RAG applications (which most production AI applications require), Azure AI Search is a significant cost component.

Tier Monthly Cost (AUD approx) Storage Indexes Best For
Free $0 50 MB 3 Experimentation only
Basic ~$110/month 2 GB 15 Small production workloads
Standard S1 ~$370/month 25 GB per partition 50 Most production workloads
Standard S2 ~$1,480/month 100 GB per partition 200 Large document collections
Standard S3 ~$2,960/month 200 GB per partition 200 Very large collections

What most mid-market Australian organisations need: Standard S1 handles most initial production deployments. You'll likely spend $370-$740 AUD/month on AI Search once you're running a production RAG application.

Vector search adds cost: If you enable vector search (recommended for better retrieval quality), you'll also pay for the embedding model used to vectorise your documents. Using text-embedding-3-small costs roughly $0.02 USD per million tokens - usually negligible compared to other costs.

Storage Costs

Azure Blob Storage for your training data, documents, and model artifacts:

  • Hot tier: ~$0.0264 AUD per GB per month
  • Cool tier: ~$0.0132 AUD per GB per month

For most AI projects, storage costs are a rounding error - typically under $50 AUD/month unless you're working with very large document collections or media files.

Networking Costs

Usually small, but worth being aware of:

  • Private endpoints: ~$13 AUD/month per endpoint. If you have 5 private endpoints for your AI Foundry setup, that's $65 AUD/month.
  • Data egress: Charges apply when data leaves Azure or moves between regions. For AI workloads staying within a single region, this is minimal.

Real-World Cost Examples

Here are actual monthly cost breakdowns from Australian organisations we work with (anonymised):

Example 1 - Internal Knowledge Assistant (Mid-Size Professional Services)

Component Monthly Cost (AUD)
GPT-4o mini inference (50,000 queries/month) $180
Azure AI Search (Standard S1) $370
Storage $15
Application Insights monitoring $30
Total $595

Example 2 - Document Processing Pipeline (Financial Services)

Component Monthly Cost (AUD)
GPT-4o inference (20,000 documents/month) $2,400
GPT-4o mini for pre-classification $85
Azure AI Search (Standard S1) $370
Fine-tuned model hosting $320
Storage (large document archive) $90
Monitoring $45
Total $3,310

Example 3 - Customer Service AI (Retail)

Component Monthly Cost (AUD)
GPT-4o mini inference (200,000 conversations/month) $720
Azure AI Search (Standard S1) $370
Content safety API calls $110
Storage $20
Monitoring $30
Total $1,250

Example 4 - Multi-Model Enterprise Platform (Large Enterprise)

Component Monthly Cost (AUD)
GPT-4o inference (complex analysis tasks) $8,500
GPT-4o mini inference (high-volume tasks) $1,800
Llama 3.1 70B on managed compute $4,000
Azure AI Search (Standard S2, 2 partitions) $2,960
Fine-tuning runs (periodic retraining) $200
Storage $150
Networking (private endpoints) $130
Monitoring $120
Total $17,860

Cost Management Strategies That Actually Work

1. Model Routing - Match the Model to the Task

This is the single highest-impact cost optimisation. Not every query needs your most expensive model.

Build a routing layer that sends queries to different models based on complexity:

  • Simple factual questions and classifications go to GPT-4o mini (or Phi-3)
  • Complex analysis, reasoning, and generation go to GPT-4o
  • Straightforward extraction tasks go to the cheapest model that meets accuracy thresholds

One client reduced monthly inference costs from $12,000 to $4,800 AUD by implementing model routing. The approach takes about 2-3 weeks to set up properly.

2. Prompt Optimisation - Reduce Token Count

Every token costs money. Optimise your prompts:

  • Remove redundant instructions (the model doesn't need to be told the same thing three ways)
  • Use concise system prompts (200-400 tokens is usually sufficient)
  • Ask for concise outputs explicitly ("respond in 2-3 sentences" vs. letting the model write an essay)
  • Reduce retrieved context in RAG applications (return 3 relevant chunks instead of 10)

A 30% reduction in average prompt length translates directly to a 30% reduction in input token costs. We've achieved this on multiple client projects just by auditing and trimming prompts.

3. Caching - Don't Pay Twice for the Same Answer

If your application receives similar queries repeatedly, implement a caching layer:

  • Cache exact-match queries and their responses
  • Use semantic similarity caching for near-duplicate queries
  • Set appropriate cache expiry based on how often your data changes

For applications with repetitive query patterns (FAQ-style, standard report generation), caching can reduce API calls by 40-60%.

4. Azure Budgets and Alerts

Set up Azure Cost Management budgets from day one:

  1. Create a budget for each AI Foundry project
  2. Set alerts at 50%, 75%, and 90% of expected monthly spend
  3. For development environments, set a hard spending cap

How to set this up: In the Azure Portal, go to Cost Management + Billing, create a budget scoped to your AI Foundry resource group, and configure action groups for email alerts.

5. Reserved Instances and Commitments

If your AI Search or managed compute costs are significant and predictable:

  • Azure Reservations: Commit to 1- or 3-year terms for 20-40% savings on compute
  • PTU commitments: Provisioned Throughput Units for high-volume inference workloads
  • Azure Savings Plans: Flexible compute commitments that apply across VM types

For an organisation spending $5,000+ AUD/month consistently, a 1-year reservation typically saves $12,000-$24,000 AUD annually.

6. Development vs. Production Environments

Don't run development at production scale:

  • Use the free tier of AI Search for development
  • Use GPT-4o mini (or even Phi-3) for development and testing, even if production uses GPT-4o
  • Shut down managed compute deployments outside business hours in development
  • Use smaller datasets for development fine-tuning runs

We've seen development environments that cost more than production because nobody thought to scale them down. A simple scheduled script that stops development compute at 6pm and starts it at 8am can save 60% of development compute costs.

7. Monitor and Optimise Continuously

AI costs change as usage patterns evolve. Set up a monthly review:

  • Review Azure Cost Management reports by resource and tag
  • Identify the top 3 cost drivers
  • Check if model routing is working as expected
  • Look for anomalies (sudden cost spikes, unexpected usage patterns)
  • Review whether cheaper models have become available since your last assessment

Microsoft regularly releases new models and pricing changes. GPT-4o mini didn't exist two years ago - organisations that were paying for GPT-4 Turbo for simple tasks could have saved 90% by switching.

Budgeting for Your First Azure AI Foundry Project

If you're planning your first project and need a budget estimate, here's a framework:

Proof of Concept (4-8 weeks)

Item Estimated Cost (AUD)
Azure AI Foundry consumption (inference, search, storage) $1,500 - $4,000
Consulting/development (if external) $15,000 - $40,000
Total $16,500 - $44,000

Production MVP (first 6 months)

Item Monthly Cost (AUD)
Inference $200 - $5,000
AI Search $370 - $1,500
Storage and networking $50 - $200
Monitoring $30 - $100
Monthly total $650 - $6,800

Ongoing Optimisation Savings

Based on our experience, a focused cost optimisation effort after the first 3 months of production typically reduces monthly spend by 25-40%. The investment in optimisation (usually 20-40 hours of work) pays for itself within 2-3 months.

Common Pricing Mistakes to Avoid

Using GPT-4o for everything: The most expensive model is not always the best model for the job. Classify your tasks by complexity and route accordingly.

Leaving development resources running: Managed compute, AI Search instances, and other resources charge 24/7 unless you actively stop or downgrade them.

Ignoring token counts in prompts: A system prompt with 2,000 tokens that runs on every API call adds up quickly at scale. Audit your prompts.

Not setting budgets: Without budgets and alerts, you won't know there's a problem until the invoice arrives.

Over-provisioning AI Search: Starting with Standard S2 when Basic would suffice for your document volume wastes hundreds of dollars per month.

Skipping the caching layer: If even 20% of your queries are repeated or nearly identical, caching pays for itself immediately.

How Team 400 Helps With Cost Management

We include cost analysis and optimisation in every Azure AI Foundry engagement. When we build AI applications for clients, we design the architecture with cost efficiency in mind from the start - model routing, prompt optimisation, caching, and appropriate resource sizing.

For organisations already running Azure AI Foundry workloads, we offer focused cost optimisation assessments. We review your current architecture, identify savings opportunities, and implement changes. Typical engagement length is 2-3 weeks, and the savings typically exceed the engagement cost within 2-3 months.

Want to understand what your Azure AI Foundry project will cost, or need help reducing your current spend? Get in touch and we'll give you an honest assessment.

Explore our Azure AI Foundry consulting, our broader AI consulting services, or learn about our approach as Microsoft AI consultants.