Does a deployed fine-tuned model cost money when nobody uses it?

Yes. A fine-tuned deployment accrues hosting charges for as long as it is deployed. Azure deletes a deployment that gets no calls for 15 days, but the model remains and can be redeployed; deleting unused deployments sooner avoids paying for that idle period.

Which token settings should we review first on a client's apps?

Start with output length: Microsoft advises setting max_tokens as low as the task allows and not requesting multiple completions per call. Then compare prompt and completion token metrics across deployments to spot oversized prompts.

Can we set a hard spending cap on a client's Azure OpenAI usage?

Not natively. Microsoft states Azure OpenAI doesn't currently offer hard limits that block spend beyond a budget. Create Cost Management budgets and alerts for each client's subscriptions or resource groups instead.

Coverage

Azure OpenAI and AI services cost optimisation for MSPs

Azure AI costs start as soon as resources are deployed: models bill for every input and output token, fine-tuned models carry an hourly hosting fee even when unused, and Azure OpenAI has no hard spending limit to stop an overrun. Without regular reviews, idle deployments and oversized prompts quietly inflate each client's bill.

18

automated finders for Azure OpenAI and AI services

Read-only, across every connected client tenant, tied to real billing data.

Modernise
Remove unused
Rightsizing
Scaling

Where Azure OpenAI and AI services spend goes to waste

The main cost levers Microsoft documents for Azure OpenAI and AI services, and the ones an MSP should check across every client.

Reserve PTUs only for steady demand

Standard pay-per-token deployments suit development, low volumes and unpredictable traffic. Provisioned throughput is billed per PTU per hour whether or not requests arrive, so it only pays off for steady, high-volume production workloads, where a reservation can lower the effective rate. ^[2]

Default to Global Standard, batch what can wait

Microsoft recommends Global Standard as the default for most workloads because it has the lowest price, using Data Zone or regional types only when residency requires it. Work that is not time-sensitive can go to Global Batch at 50% less than Global Standard, with a 24-hour target turnaround. ^[3]

Use smaller models where they do the job

Not every request needs a large or reasoning model. Foundry's model router picks a model for each prompt behind a single deployment, and its cost mode favours cheaper models for high-volume, budget-sensitive workloads. Benchmark it against your current baseline first. ^[4]

Structure prompts for cache hits

Prompt caching is on by default for supported models, and cached input tokens are billed at a discount. A prompt needs at least 1,024 tokens with an identical opening to hit the cache, so put stable instructions and tool definitions first and variable content last. ^[5]

Remove idle deployments billed by the hour

Fine-tuned models on Standard deployments add an hourly hosting fee on top of token charges, and provisioned or managed-compute deployments bill for allocated capacity with no traffic. Test fine-tuned models on the Developer deployment type, which has no hosting fee. ^[6]

Move steady AI services usage to commitment tiers

Speech, translation, Language, Vision and Document Intelligence offer commitment tiers: a fixed monthly fee at a discount to standard pricing, with overage above the included quota. They can't be refunded once bought, so match them to proven, steady volumes. ^[7]

How Cerberus Cloud covers Azure OpenAI and AI services

Cerberus Cloud runs 18 automated finders for Azure OpenAI and AI services across every connected client tenant. Each one ties a technical signal to real billing data, so findings arrive with an estimated recoverable value, ranked by impact, effort and risk, with evidence and remediation scripts for your engineers to review.

Access is read-only, so nothing changes in a client environment until your team decides it should. See how it works or the multi-tenant portfolio view.

Modernise
Remove unused
Rightsizing
Scaling

Frequently asked questions

Does a deployed fine-tuned model cost money when nobody uses it?: Yes. A fine-tuned deployment accrues hosting charges for as long as it is deployed. Azure deletes a deployment that gets no calls for 15 days, but the model remains and can be redeployed; deleting unused deployments sooner avoids paying for that idle period.
Which token settings should we review first on a client's apps?: Start with output length: Microsoft advises setting max_tokens as low as the task allows and not requesting multiple completions per call. Then compare prompt and completion token metrics across deployments to spot oversized prompts.
Can we set a hard spending cap on a client's Azure OpenAI usage?: Not natively. Microsoft states Azure OpenAI doesn't currently offer hard limits that block spend beyond a budget. Create Cost Management budgets and alerts for each client's subscriptions or resource groups instead.

See Azure OpenAI and AI services findings on a sample portfolio

A 15-minute walkthrough: tenant connection, prioritised findings, remediation scripts and savings history.

Book a Walkthrough All coverage

Sources

Last checked 15 October 2026. Product details change; tell us if something here is out of date.

All 15 coverage families