What is Foundry Tools line item in Azure Cost Analysis?

Question

What is Foundry Tools line item in Azure Cost Analysis? This has got the biggest footprint in my billing and being charged separately from the Foundry Models.

Answer 1

Hello bamb00,

Welcome to Microsoft Q&A and Thank you for reaching out.

In addition to Evaluations and Observability token usage, it’s important to understand that “Foundry Tools” is an umbrella billing category that can include several platform-level capabilities beyond evaluation workflows.

Even if you are not intentionally running benchmarking or monitoring jobs, Foundry Tools charges can also come from:

1.Agent Tool Orchestration Runtime

If you are using Agents in Azure AI Foundry, the platform performs:

Tool planning and routing

Function/tool invocation orchestration

Conversation state management

• Retrieval pipelines

Even when the underlying model inference is billed elsewhere, the orchestration layer and tool execution compute are billed under Foundry Tools.

In agent-heavy workloads, this orchestration cost can exceed direct model inference costs.

2.Knowledge & Retrieval Pipelines

If your project uses Knowledge bases, Vector search / hybrid retrieval, Grounding connections, Fabric / SQL / Search connectors, MCP servers (preview)

Each request may trigger Retrieval calls, Context assembly, Additional model calls

These supporting operations are typically metered under Foundry Tools rather than Foundry Models.

3.Prompt Flow & Development Runs

Beyond formal “Evaluations,” development activities can also drive usage:

Prompt Flow test runs

Batch testing pipelines

Playground experimentation

CI/CD validation flows

Frequent test iterations can generate significant token consumption that appears under Foundry Tools.

4.Tool-Enabled Model Calls vs Direct Deployments

Another important distinction:

If you call a model through a direct Azure OpenAI deployment endpoint, costs appear under model deployment meters.

If you call the same model through Foundry SaaS endpoints, Agents, or evaluation pipelines, usage may appear under Foundry Tools.

So the billing category depends on the access path, not just the model itself.

How to Isolate Non-Evaluation Tool Costs

Since Evaluations/Observability has already been identified in your screenshot, you can further validate:

In Cost Analysis:

• Group by Meter
• Then group by Operation name

Look for non-evaluation meters (e.g., orchestration or tool execution related).

Compare timestamps of spikes with:

• Agent usage
• Prompt flow activity
• Scheduled runs

This will help determine whether the driver is:

• Monitoring workloads
• Agent/tool orchestration
• Development testing
• Or intentional benchmarking

Next Steps

If the footprint is larger than expected:

• Review whether continuous evaluation is enabled
• Check for scheduled or recurring evaluation jobs
• Review agent tool configurations (especially automatic retrieval)
• Reduce dataset sizes for monitoring runs
• Lower sampling rates
• Remove unused schedules or dev projects

Key Takeaway

“Foundry Tools” is not just model tokens, it represents platform-level intelligence, orchestration, and monitoring layers that operate on top of your models.

If it’s your largest cost category, it usually means one of two things:

• You are running evaluation/monitoring workflows at scale
• Or you are heavily using agents, retrieval, or tool-enabled pipelines.

Please refer this, Plan and manage costs for Microsoft Foundry https://learn.microsoft.com/azure/ai-foundry/concepts/manage-costs?view=foundry-classic

I Hope this helps. Do let me know if you have any further queries.

---

If this answers your query, please do click Accept Answer and Yes for was this answer helpful.

Thank you!

Answer 2

“Foundry Tools” in Azure Cost Analysis refers to Azure AI Foundry platform features that run on top of your models, not the model inference itself. It typically includes services like Evaluations, Observability, tracing, prompt flow runs, safety evaluations, and monitoring pipelines. These tools call your deployed models behind the scenes and consume tokens, but they are billed under a separate meter category from “Foundry Models.”

In your screenshot, the meters “Evaluations input tokens” and “Evaluations output tokens” under the “Observability” tier mean that Azure AI Foundry’s evaluation or monitoring workflows are sending prompts to a model and being charged per token for both input and output. This is separate from standard application inference traffic. If you have automatic evaluations enabled, scheduled benchmarking, continuous quality monitoring, or large evaluation datasets, those runs can easily generate more tokens than your production traffic, which explains the large footprint.

Common causes include running batch evaluations on large datasets, enabling continuous monitoring that replays production prompts, automated red-teaming or safety scans, and frequent prompt flow test runs in development. Every evaluation run effectively performs model inference at scale, and all tokens consumed are billed under Foundry Tools rather than under the model deployment’s normal usage line.

To confirm the source, go to Cost Analysis and group by Meter and Resource, then filter for “Evaluations input tokens” or “Evaluations output tokens.” You can also group by Resource Group to see which Foundry project or Cognitive Services resource is generating the charges. In Azure AI Foundry, check the Evaluations and Observability sections for recent or scheduled runs and review dataset sizes and run frequency.

If this usage is unintended, disable continuous evaluations, reduce dataset size, lower sampling rates, or delete unused evaluation schedules. If it is intentional, treat it as separate token-based consumption driven by testing and monitoring workloads rather than end-user inference.

---

If the above response helps answer your question, remember to "Accept Answer" so that others in the community facing similar issues can easily find the solution. Your contribution is highly appreciated.

hth

Marcin