Back to Blog

Azure AI Foundry Pricing 2026: What Does Azure AI Foundry Cost for Business?

Henri Jung, Co-founder at Superkind
Henri Jung

Co-founder at Superkind

A vintage brass cash register with its closed drawer and the orange Superkind mascot sitting on top

Azure AI Foundry pricing is the question that lands on a CTO’s desk the week a business unit asks for a custom AI agent. Microsoft lists the platform as free to use and then meters almost everything you build on it: model tokens from 11,000+ models in Foundry Models (GPT-4o at about $2.50 per million input tokens and $10 per million output, GPT-4.1 near $2 and $8, Claude Sonnet 4.5 at $3 and $15, Mistral and Meta cheaper), Foundry Agent Service tools (file search per GB per day with the first 1 GB free, code interpreter per session, web search per 1,000 transactions), hosted agent compute per vCPU and GiB hour, Azure AI Search for retrieval, and reserved capacity (PTUs) for steady loads from roughly $1 per PTU per hour. This article puts every published Foundry price in one place as of November 2026, works out what a small team, a mid-sized rollout and a production load really cost per month, and shows how Superkind’s ready-made AI team for IT and platform work does the same work from EUR 50 per month for the whole company with unlimited users, without an Azure subscription.

Key takeaways

Platform free, consumption billed: Azure AI Foundry has no platform fee; you pay for Foundry Models tokens, Foundry Agent Service tools, hosted agent compute, Foundry IQ (Azure AI Search), fine-tuning and anything else you wire in12.

Model tokens dominate the bill: GPT-4o around $2.50 / $10 per million input/output tokens, GPT-4.1 around $2 / $8, Claude Sonnet 4.5 around $3 / $15, Mistral Large near $0.50 / $1.50, with Batch at 50 percent off Global Standard and PTU reservations for steady loads3. Every model has its own rate card; a multi-vendor bill is the norm.

Get a ready-made AI team instead: Superkind comes with an IT and platform team (Cloud Spend Watcher, Knowledge Finder, Quota Requester, Model Scout) that runs automatically in Outlook, Teams, Slack and the web app. Start for free with unlimited users, or build your own on the Superkind platform.

Keep control, keep data in Europe: Superkind is hosted in Europe, GDPR compliant, aligned with the EU AI Act, no training on your data, with 1,500+ integrations including SharePoint, Excel, Outlook, Personio, HubSpot and DATEV.

Your AI team for IT and platform work

#platform-spend
Cloud Spend WatcherAI employee
Routine · every Monday at 07:30
Week 45 Azure and Foundry spend. Three lines over the 20% alert: Azure OpenAI +41%, Foundry Agent Service +28%, Azure AI Search +22%. Three owners tagged, details in the thread.
Azure OpenAI+41% vs. last week · Tim
Foundry Agent Service+28% vs. last week · Anna
Azure AI Search+22% vs. last week · Jana
Runs automatically every Monday
Example: Cloud Spend Watcher at work in Slack

What is Azure AI Foundry pricing?

Azure AI Foundry pricing is the consumption-based set of per-token, per-session, per-transaction and per-compute-hour rates Microsoft charges for the services inside Microsoft Foundry (formerly Azure AI Foundry), built on an Azure subscription. The platform itself is free to use and explore; every service inside it has its own billing model: Foundry Models bills per million input and output tokens per model, Foundry Agent Service bills per session, per GB per day, per 1,000 transactions and per compute hour for hosted agents, Foundry IQ uses Azure AI Search, and reservations (PTUs and Agent Commit Units) buy capacity in advance for a discount12. One example of a different pricing model is Superkind: a ready-made AI team priced per company with unlimited users, hosted in Europe, that does IT and operations tasks across Outlook, Teams, Slack and SharePoint through 1,500+ integrations.

How it works with Superkind: your AI team for IT and platform work

Azure AI Foundry is a build-your-own-agent platform on the Microsoft Azure stack. If your IT and platform team just wants real work done in email, Teams, Slack and the web app without wiring up Foundry Models, Agent Service meters and Azure AI Search, there is a simpler route. Superkind is the AI employee for IT and platform teams: your CTO, platform lead, help-desk and ops people write to Superkind in Outlook, Microsoft Teams, Slack or the Superkind web app like you would to a colleague, and it does the task in your programs through 1,500+ integrations such as SharePoint, Excel, Outlook, Personio, HubSpot and DATEV. Superkind is hosted in Europe, GDPR compliant, aligned with the EU AI Act and never trains models on your data. It is priced per company with unlimited users, not per seat and not per token.

You do not start from scratch: Superkind comes with ready-made AI employees for IT and platform teams. They run on their own in the background and you decide where they ask you first. Use them right away, adapt them, or build your own on the platform.

Cloud Spend Watcher

AI employee for cloud cost reviews

Nobody has time to open the Azure, Microsoft 365 and OpenAI bills every week to spot the line that moved. Every Monday morning the Cloud Spend Watcher pulls the latest Azure invoices and token-usage reports, compares them with last month in the Excel sheet in SharePoint and posts a short summary in the Slack platform channel. Any line that moves by more than 20 percent is flagged with the owner for approval, the rest keeps running.

Works inSlackExcelSharePoint+ 1,700 more
#platform-spend
Cloud Spend WatcherAI employee
Routine · every Monday at 07:30
Week 45 Azure and Foundry spend. Three lines over the 20% alert: Azure OpenAI +41%, Foundry Agent Service +28%, Azure AI Search +22%. Three owners tagged, details in the thread.
Azure OpenAI+41% vs. last week · Tim
Foundry Agent Service+28% vs. last week · Anna
Azure AI Search+22% vs. last week · Jana
Runs automatically every Monday
Example: Cloud Spend Watcher at work in Slack

Knowledge Finder

AI employee for internal platform knowledge

Every day the platform lead gets the same questions in Teams: which region to deploy in, which model to use for a new prototype, how to request a Foundry quota. The Knowledge Finder watches the #platform-help channel, finds the answer in SharePoint and the platform docs, posts it with a direct link, and marks the thread as resolved. Routine lookups run on their own; new decisions are added by the lead once and the whole team sees them next time they ask.

Works inMicrosoft TeamsSharePointWord+ 1,700 more
platform-helpLinnartz GmbH
Markus Lehmann, Backend
Which Foundry region should I pick for a new agent that handles German customer data? And which model is the default for a prompt-only prototype?
Knowledge FinderAI employee
New message in #platform-help
Default region for German customer data is Sweden Central with the EU Data Zone. Default model for a prompt-only prototype is GPT-4.1 mini on Global Standard. Full policy and the request form in the links below.
PolicyPlatform guide · Oct 2026
SourceSharePoint · Platform Docs
Request formlinked
Runs on its own on every new question
Example: Knowledge Finder at work in Microsoft Teams

Quota Requester

AI employee for Azure quota and access

Every new Azure OpenAI or Foundry quota request lands in the IT inbox with a form, a region, a model and a business reason. The Quota Requester picks up each new email on its own, checks the requester in SharePoint, fills the internal quota form with region and model, drafts the Azure support request and shows the full package before anything is submitted. Only the final submit waits for your OK.

Works inOutlookSharePointMicrosoft Teams+ 1,700 more
Foundry quota: GPT-4.1 for the new recommender
Lena Fischer, Data Platform
Hi IT, I need 20 PTUs for GPT-4.1 on Global Provisioned for the new recommender service, Sweden Central, use case is product recommendations for our German webshop. Can you file the Azure request?
Quota RequesterAI employee
New email to it@
Request prepared. Model GPT-4.1, Global Provisioned, 20 PTUs, Sweden Central, EU Data Zone, requester Lena Fischer (Data Platform), budget code DP-2026-11. Waiting for your OK before I submit to Azure support.
ModelGPT-4.1 Global Provisioned
RegionSweden Central · EU Data Zone
PTUs20
Submit the Foundry quota request to Azure support?Hi Azure support, we would like to request 20 PTUs for GPT-4.1 Global Provisioned in Sweden Central for an internal recommender service ...ApproveReject
Example: Quota Requester at work in Outlook

Model Scout

AI employee for model catalogue tracking

Foundry Models adds new models almost every week and the platform lead has to decide whether the team should try them. On the first working day of each month the Model Scout pulls the current Foundry Models catalogue and the token-usage report from Azure, writes a short review (new models worth testing, models no one used, candidates to retire) and posts the draft in the Superkind app for the lead. One click turns the draft into a Teams post for the engineering channel.

Works inMicrosoft TeamsSharePointExcel+ 1,700 more
Superkind
New chat
My tasks
AI employees
Model Scout
Cloud Spend Watcher
Knowledge Finder
Monthly model review · November 2026

On the first working day of each month: pull the Foundry Models catalogue and the token-usage report, draft a short monthly model review (worth testing, unused, to retire), show it for one-click approval before posting to the engineering channel.

Reasoning ⌄
Model Scout
Routine · first working day of the month

November review drafted. Worth testing: two new Mistral reasoning models and one Claude Haiku 4.5 variant. Unused last month: 4 models (0 tokens). Candidates to retire: 2 models with less than 0.1% of monthly token spend.

Worth testing3 models
Unused4 models
To retire2 candidates
Post the November model review to the engineering channel?Hi team, here is the November Foundry Models review. Three new candidates worth a two-hour spike ...
ApproveReject
Send a message ...
Example: Model Scout at work in Superkind

Every employee is ready to use right away. Or build your own on the platform.

Safe by default

Superkind is hosted and processed in Europe, GDPR compliant, aligned with the EU AI Act and does not train models on your data. The AI models come from OpenAI, Anthropic, Mistral and Google, all processed in Europe. A data processing agreement is available through the Trust Center at trust.superkind.ai. For IT and platform work that means cloud invoices, platform docs, quota requests and model-usage reports stay in your Europe-hosted setup.

Send Superkind your first platform task

Start for free in Outlook, Teams or Slack. No Azure subscription needed.

From the first task to your own app

With Superkind the price grows with the work, not with the number of licences or with every token drawn. Four stages, on the IT and platform example:

  1. 1

    Ready in minutes

    Sign up for free at app.superkind.ai with up to EUR 150 in credits, forward one real IT task from the it@ mailbox to Superkind in Outlook, Teams or Slack and watch it get done. No Azure subscription, no Foundry Models deployment, no partner project.

  2. 2

    Tasks become routines

    What you explain once keeps running. Monday Azure and Foundry cost review in Slack, platform questions answered in the Teams help channel, every new Foundry quota request drafted from the IT inbox. You decide which steps wait for approval, such as submitting a quota request or spinning up a new PTU reservation.

  3. 3

    Learns from day one

    Every task makes Superkind smarter. Your Azure region policy, the default model per use case, the owners for every system, the Foundry quota playbook end up in your Company Brain, and that knowledge stays with the team even when a platform engineer leaves.

  4. 4

    Your own app, without starting over

    When platform work gets more complex, you don’t start over. Build custom solutions, apps and integrations on the Superkind platform, on top of what Superkind already knows about your estate, for example your own model-review workbench or a self-service quota portal. Your Super App. Built yourself or with us. It all belongs to you.

A tray of blank unmarked round graphite coins with one small coral-orange accent on the rim of a single coin

How Azure AI Foundry pricing works

Microsoft sells Foundry as an Azure workload with one platform (free) and several meters that bill in parallel. Most teams end up with three to five of these meters running at once.

  1. Azure subscription - required to deploy any Foundry resource. A new Azure account comes with $200 in credit for 30 days to try services out9.
  2. Foundry Models (tokens) - every answer draws from a model’s token rate. Pay-as-you-go Standard per million input and output tokens; Batch API is 50 percent off Global Standard3.
  3. Foundry Agent Service (tools) - running a prompt or workflow agent has no extra charge, but file search bills per GB per day (first 1 GB free), code interpreter per session, web search and custom search per 1,000 transactions, memory per GiB-hour for long-term and short-term, and hosted agents per vCPU hour and per GiB hour on the underlying container2.
  4. Foundry IQ and Azure AI Search - retrieval over your own documents uses Azure AI Search, billed as its own Azure service (replicas, partitions, storage).
  5. Reserved capacity (PTUs) - instead of paying per token, you reserve Provisioned Throughput Units per model and region, billed per hour, with monthly and yearly commitments at a discount. As of November 2026, third-party guidance puts Global Provisioned near $1 per PTU per hour and reservation discounts at roughly 20-40% for one month and 40-65% for one year10. Microsoft does not publish a fixed PTU price; ask for a written quote.
  6. Agent Commit Units (ACUs) - a one-year pre-purchase of units used across included Microsoft services, at three tiers: Tier 1 (20,000 ACUs, 5% discount), Tier 2 (100,000 ACUs, 10%), Tier 3 (500,000 ACUs, 15%)1.
  7. Fine-tuning - training hours plus per-hour hosting of the fine-tuned model deployment, on top of its own per-token input and output rates3.
  8. Content Safety and governance - Content Safety for text, images and prompt-shield checks is billed per 1,000 records, usually a small line but it is its own meter.

Two commitments decide most of the bill: which model you default to (GPT-4.1 mini and Mistral are a fraction of GPT-4o or Claude Opus), and whether your load is steady enough to justify PTUs instead of Standard. The third big lever is region: an EU Data Zone deployment in Sweden Central or Germany West Central keeps data in Europe and typically prices slightly above Global Standard7.

The 6 Azure AI Foundry pricing meters

This table lists the main Foundry meters, with Superkind in the first row for comparison. Foundry rates are from the Microsoft Azure AI Foundry, Foundry Agent Service and Azure OpenAI pricing pages (November 2026); Superkind rates are from Superkind’s public pricing (Stand: Oktober 2026). Prices excl. VAT, Microsoft in USD.

MeterPriceBilling unitWhat you get
Superkind (for comparison)Free: EUR 0 with a one-off up to EUR 150 in credits
Business: from EUR 50 / month (25,000 credits), EUR 100 for 50,000 credits
Enterprise: custom
Per company, unlimited users, cancel monthlyAn AI employee in Outlook, Teams, Slack and the web app that does platform and operations tasks in your programs through 1,500+ integrations (SharePoint, Excel, Outlook, Personio, HubSpot, DATEV ...), with approval, hosted in Europe
Foundry platform$0Included with an Azure subscriptionPortal access, project management, agent builder, free to use and explore1
Foundry Models (Standard)Per model, per million tokens (e.g. GPT-4o ~ $2.50 in / $10 out, GPT-4.1 ~ $2 / $8, Claude 4.5 Sonnet $3 / $15, Mistral Large ~ $0.50 / $1.50)3Per million input and output tokensPay-as-you-go access to 11,000+ models across OpenAI, Anthropic, Mistral, Meta, Microsoft and more4
Foundry Agent Service (tools)File search per GB / day (1 GB free), code interpreter per session, web search and custom search per 1,000 transactions, memory per GiB-hour, hosted agents per vCPU hour and per GiB hour2Per session, per GB-day, per 1,000 transactions, per compute hourAgent runtime, file search, code interpreter, web search, memory and hosted container runtime for agents2
Reserved capacity (PTUs)From ~ $1 / PTU / hour on Global Provisioned (third-party guidance), 20-40% off on monthly reservations, 40-65% off on yearly10Per PTU per hour; minimum 15 PTUs per model on Global, 50 on RegionalReserved throughput for OpenAI, Microsoft Phi and DeepSeek models; third-party models stay pay-as-you-go6
Agent Commit Units (ACUs)Tier 1: 20,000 ACUs, 5% discount
Tier 2: 100,000 ACUs, 10% discount
Tier 3: 500,000 ACUs, 15% discount1
Per year, bought in the Azure portalA pre-purchased pool of units usable across included Microsoft services inside Foundry1
Fine-tuningPer training hour + per hour of hosting the fine-tuned deployment + its own input/output token rates3Per hour training, per hour hosting, per million tokensSupervised and reinforcement fine-tuning on GPT-4o, GPT-4.1 and o4-mini families3

Beyond these meters, the model rate card inside Foundry Models shapes every quote. Picking GPT-4.1 mini over GPT-4o cuts the per-token bill by roughly 90 percent at Standard rates; picking Mistral Large or an open-source Meta Llama over Claude Opus 4 can cut it by 95 percent, at the cost of different model behaviour that the team has to validate.

Model rate card: GPT-4o, GPT-4.1, o-series, Claude, Mistral

The headline rates for the models most teams default to, in USD per million tokens on Global Standard pay-as-you-go. These are widely cited third-party aggregator figures that line up with the Azure OpenAI and Foundry Models documentation as of November 2026; the official pricing page renders the exact numbers as placeholders and expects you to check the live page or the Azure pricing calculator before you commit34. Batch API is 50 percent off Standard, Data Zone and Regional run slightly above Global.

Model (Global Standard)Input / 1M tokensOutput / 1M tokensTypical use
GPT-4o~ $2.50~ $10.00Strong default chat and multimodal model
GPT-4o mini~ $0.15~ $0.60High-volume classification, lightweight chat
GPT-4.1~ $2.00~ $8.00Long-context reasoning, coding, tool use
GPT-4.1 mini~ $0.40~ $1.60Cheap long-context default for prototypes
o3~ $2.00~ $8.00Deep reasoning on hard problems
o3-mini~ $1.10~ $4.40Cheaper reasoning for batch jobs
Claude Sonnet 4.5~ $3.00~ $15.00Long-context writing and agent work
Claude Haiku 4.5~ $0.80~ $4.00Cheap Claude-family model
Claude Opus 4~ $15.00~ $75.00Top-of-range Claude for the hardest tasks
Mistral Large~ $0.50 - $0.63~ $1.50 - $1.88European model, strong general-purpose
Mistral Small~ $1.00 (or lower on light variants)~ $3.00Cheaper Mistral tier

A couple of practical rules of thumb. A typical business chat turn (300 input, 400 output tokens) on GPT-4o costs about $0.0048, or $0.48 per 100 chats; on GPT-4.1 mini the same turn is about $0.0007, or $0.07 per 100 chats. A long-document RAG answer (8,000 input, 1,000 output tokens) on Claude Sonnet 4.5 is about $0.039 per answer, or $39 per 1,000 answers. Multiply by your real daily volume to size a Foundry bill.

Examples: what teams pay for Azure AI Foundry

Five realistic set-ups. All figures are calculated from the published or widely cited rates above, excl. VAT, and excl. Microsoft 365 or Azure base licences unless stated. We use dollar list prices on Azure and treat the mid-point of ranges as the estimate.

1. The IT and platform team runs on Superkind

Tim’s IT and platform team of 10 writes to Superkind in Outlook, Teams and Slack. Superkind posts the Monday Azure and Foundry cost review in Slack, answers platform questions in the Teams help channel, drafts every new Foundry quota request from the IT inbox and ships the monthly Foundry Models review. As of October 2026 you start for free with a one-off up to EUR 150 in credits; Business costs from EUR 50 per month for the whole company with unlimited users, and EUR 100 covers up to 1,000 tasks or 5,000 questions.

2. Small-team prototype on Foundry (GPT-4.1 mini)

A 10-person platform team runs a GPT-4.1 mini prompt agent in the Foundry portal for an internal knowledge tool: 20,000 requests per month at 800 input + 400 output tokens each. Tokens: 20,000 x 800 x $0.40 / 1M input + 20,000 x 400 x $1.60 / 1M output = $6.40 + $12.80 = about $19 per month on tokens. Add file search (3 GB over the free 1 GB) at a small per-GB-per-day rate and a handful of code-interpreter sessions, and the whole prototype lands under $100 per month23. Fine for a pilot; the bill grows with traffic and with a jump to GPT-4o.

3. Production customer-facing agent on GPT-4o

A B2B retailer runs a customer-facing Foundry agent on its website. 3,000 chats per month at 1,500 input + 800 output tokens each, plus one web search per chat. Tokens: 3,000 x 1,500 x $2.50 / 1M + 3,000 x 800 x $10 / 1M = $11.25 + $24 = about $35 per month on tokens. Add 3,000 web-search transactions per month (billed per 1,000) and a 10 GB vector-storage pool for file search, plus an Azure AI Search tier for Foundry IQ, and the total usually lands between $500 and $1,500 per month at Standard rates23. Grow to 30,000 chats and token spend becomes the headline.

4. Reserved capacity with PTUs for a steady internal agent

A 500-person company runs a GPT-4.1 internal agent in Teams all day, averaging 100 requests per minute across working hours and well above 50 percent utilisation of its throughput. The platform team reserves 15 PTUs of GPT-4.1 Global Provisioned (the model minimum)3. At the third-party guidance of about $1 per PTU per hour: 15 x 24 x 30 = 10,800 PTU-hours = $10,800 per month at hourly; a one-month reservation (say 30 percent discount) lands near $7,500 per month; a one-year reservation (50 percent discount) near $5,400 per month10. Microsoft publishes no fixed PTU rate - ask for a written quote.

5. Multi-model agent on Claude and Mistral

A legal team builds a Foundry agent that uses Claude Sonnet 4.5 for drafting and Mistral Large for cheap triage. 2,000 drafts per month at 6,000 input + 1,500 output tokens on Claude: 2,000 x 6,000 x $3 / 1M + 2,000 x 1,500 x $15 / 1M = $36 + $45 = $81 per month on Claude tokens. 10,000 triage calls at 500 input + 200 output tokens on Mistral Large: 10,000 x 500 x $0.55 / 1M + 10,000 x 200 x $1.70 / 1M = $2.75 + $3.40 = about $6 per month on Mistral tokens3. Note that Anthropic and Mistral models were not reservation-eligible via PTUs at the time of writing (only OpenAI, Microsoft Phi and DeepSeek were)6.

Teams of 10 and 50: Foundry vs. Superkind

Plain arithmetic for a team of 10 people and a team of 50 people running the same workload of 500 and 2,500 agent requests per month respectively, each using roughly 1,500 input + 800 output tokens on GPT-4o, one file search over a 5 GB vector store and one web search per request. All figures per month, excl. VAT, Foundry in USD, Superkind in EUR. Azure AI Search, Azure storage and partner implementation are not included; see the next section.

Set-upTeam of 10 (500 req/month)Team of 50 (2,500 req/month)
Superkind Business (whole company, unlimited users, EU)EUR 50-100 / month
(EUR 100 covers 1,000 tasks or 5,000 questions)
EUR 100-200 / month
(top up when more work comes in)
Foundry Standard on GPT-4.1 miniTokens ~ $0.80 + tools ~ $30
~ $30-60 / month, ~ $360-720 / year
Tokens ~ $4 + tools ~ $100
~ $100-200 / month, ~ $1,200-2,400 / year
Foundry Standard on GPT-4oTokens ~ $6 + tools ~ $50
~ $100-200 / month, ~ $1,200-2,400 / year
Tokens ~ $29 + tools ~ $200
~ $300-600 / month, ~ $3,600-7,200 / year
Foundry Standard on Claude Sonnet 4.5Tokens ~ $12 + tools ~ $50
~ $150-300 / month, ~ $1,800-3,600 / year
Tokens ~ $58 + tools ~ $200
~ $500-1,000 / month, ~ $6,000-12,000 / year
Foundry PTU reservation (only above ~50% utilisation)Rarely worth it below 15 PTUs
~ $5,400 / month minimum (1-yr on Global)
Rarely worth it below 15 PTUs
~ $5,400 / month minimum (1-yr on Global)

The two products buy different things. A Foundry token pays for one model call on the model you picked, metered input and output separately. A Superkind credit pays for a task or a question done in your programs across the team, for any number of people; EUR 100 covers up to 1,000 tasks or 5,000 questions. Many platform teams end up using both: Foundry for the custom AI apps a cloud team builds and runs end-to-end, Superkind for the day-to-day IT and operations work that crosses Outlook, Teams, Slack, SharePoint, HubSpot, Personio and DATEV without wiring new token meters for every new use case. For a wider comparison see also ChatGPT pricing, Claude pricing and Copilot Studio pricing.

What is missing and what costs extra

The pricing pages show tokens, sessions and transactions. The real quote has more lines. Five items that regularly show up on an Azure or partner invoice:

  • Azure AI Search for Foundry IQ - retrieval over your own documents runs on Azure AI Search, billed as its own service per hour per tier (replicas, partitions, semantic ranker, high-storage). A mid-sized index with 3 replicas on Basic or Standard easily adds $200-800 per month before you touch Foundry.
  • Azure storage, networking and monitoring - blob storage for files an agent reads, egress for responses, Application Insights or Log Analytics for traces. Individually small, together a visible line.
  • Content Safety and governance - Content Safety for text, images and prompt-shield checks is billed per 1,000 records; Defender for Cloud alerts and Microsoft Purview sensitivity labels are their own meters.
  • Fine-tuning and dedicated hosting - fine-tuning jobs bill per training hour, then the fine-tuned deployment bills per hour of hosting plus its own token rates. One deployment per model, so a team with three fine-tunes carries three hosting bills in parallel3.
  • Implementation and ownership - partner day rates to build agents, connectors and knowledge bases, plus the cloud engineer time to curate the Foundry IQ sources, tune prompts, write evals and keep the stack running. Both outside the Microsoft price list.

Data protection and EU hosting

Cloud invoices, platform docs, quota requests and model-usage reports often carry personal and confidential data, so whichever tool you pay for, check where it is processed, whether it is used for training and that a DPA is in place. Microsoft offers the EU Data Boundary for most Microsoft Cloud services and Foundry EU Data Zone deployments for Azure OpenAI pinned to Sweden Central and Germany West Central; Global Standard deployments can route prompts outside the chosen region, so pick Data Zone or Regional for personal data and sign the Microsoft Products and Services DPA78.

What Superkind costs

Stand: Oktober 2026. Prices excl. VAT, for the whole company with unlimited users.

  • Free - EUR 0 with a one-off up to EUR 150 in credits to start. No credit card, no sales call, every feature and every connection, in Slack, Teams, email and the app.
  • Business - from EUR 50 per month for 25,000 credits. EUR 100 for 50,000 credits is the popular step, then EUR 200 and EUR 500. No price per seat, unused credits carry over one month, top up when needed, custom apps included, cancel monthly.
  • Enterprise - custom price for your volume, dedicated support with SLA, SSO/SCIM and audit log, a DPA and help with the security review, payment by invoice.
  • What credits buy - a question uses 10-50 credits, a task 50-500 credits and a run of your own agent 1,000-5,000 credits. EUR 100 covers up to 1,000 tasks or 5,000 questions.

When Azure AI Foundry is worth it, and when it is not

Foundry is a real product with real customers. The question is fit for your team and your stack.

  • Good fit - you have an Azure cloud team, your data lives in Azure Blob, SharePoint and Azure AI Search, you want full control over models, prompts and evals, and you are building a custom AI app or API that the whole product plans to depend on. The platform’s depth pays off when you run it as a product.
  • Poor fit - you do not have an Azure team to build and run agents, your data sits in Slack, HubSpot, DATEV, Personio, Google Workspace or a mix, you want a price per company rather than per token, or you need EU residency and the DPA in place from day one for every surface without sorting out Data Zone deployments.
  • Mixed - you have an Azure team and want to try one small internal agent (a help-desk bot in Teams, a policy answer engine). Pay-as-you-go on GPT-4.1 mini or Mistral Large keeps a pilot under $100 per month, and you can promote it to GPT-4o or Claude Sonnet 4.5 once volume justifies it3.

If any of the poor-fit lines ring true, start Superkind in parallel with up to EUR 150 in credits on one real task, such as triaging the IT inbox or writing the Monday Azure cost review. Compare the Superkind result with your Foundry quote before you sign. See also our overview What is Azure AI Foundry? and the decision guide Which LLM for the Mittelstand?.

How to start

You can decide this in one afternoon. Three steps:

Your 3-step check

  • Test an AI employee on one real IT task - start Superkind for free at app.superkind.ai with up to EUR 150 in credits and send one real task by email, Teams or Slack, such as writing next Monday’s Azure cost review or drafting the latest Foundry quota request.
  • Get the full Foundry quote - ask Microsoft (or your partner) for a written quote that lists token volume per model, Foundry Agent Service tool spend, hosted agent compute, Azure AI Search tier, PTU reservations, fine-tuning hosting and partner implementation, for year one and year two.
  • Compare on the same task - put the Superkind result next to the Foundry quote on one week of real IT and platform work, not on a feature list. Pick the one that wins on the actual output.

Conclusion

Azure AI Foundry is free as a platform and paid as a stack of meters: Foundry Models tokens at $0.15 to $75 per million tokens depending on the model, Foundry Agent Service tools (file search, code interpreter, web search, memory, hosted agents), Foundry IQ (Azure AI Search), reserved capacity (PTUs) from roughly $1 per PTU per hour on Global Provisioned, Agent Commit Units, fine-tuning and Content Safety. Two commitments decide most of the bill: which model you default to, and whether your load is steady enough for PTUs. For teams in Slack, HubSpot, Personio, DATEV or a mixed stack, Superkind costs EUR 0 to start and from EUR 50 per month for the whole company with unlimited users, works through 1,500+ integrations such as SharePoint, Excel, Outlook, Personio and DATEV, and is hosted in Europe (Stand: Oktober 2026). Send Superkind one real platform task this week and put the result next to your Foundry quote.

Frequently Asked Questions

The Azure AI Foundry platform itself is free to use and explore; you pay for what you consume. As of November 2026, a small internal prototype on GPT-4.1 ($2 per million input tokens, $8 per million output tokens on Global Standard) with light file-search storage and a handful of agent sessions lands between $50 and $300 per month. A production agent on GPT-4o ($2.50 input, $10 output per million tokens) with a few thousand conversations per month and web search tools typically sits between $500 and $3,000 per month. Reserved capacity (PTUs) only pays off above roughly 50 percent utilisation and starts at 15 PTUs for GPT-4.1 and GPT-4o on Global Provisioned. Add Azure AI Search, Azure storage and partner implementation on top.

No. Azure OpenAI is the API for OpenAI models hosted on Azure, and it still exists inside Foundry Models. Azure AI Foundry (now called Microsoft Foundry since January 2026) is the broader platform: a catalogue of 11,000+ models from OpenAI, Anthropic, Mistral, Meta, Microsoft and others, the Foundry Agent Service for building and running agents, Foundry IQ for retrieval (powered by Azure AI Search), Foundry Tools and Foundry Local for edge deployment. Azure OpenAI pricing is a subset of Foundry pricing, and most new features only ship in Foundry.

A Provisioned Throughput Unit (PTU) is reserved model capacity billed per hour, with monthly and yearly commitments on top. As of November 2026 published guidance puts Global Provisioned near $1 per PTU per hour, Data Zone near $1.10 and Regional near $2, with per-model minimums of 15 PTUs for GPT-4.1 and GPT-4o on Global, 50 on Regional, 25 for the mini variants, and discounts of roughly 20-40% on a one-month reservation and 40-65% on a one-year reservation. The exact rate is not on a public price list and is set per customer; check the Azure pricing calculator and ask Microsoft for a written quote per model and region.

Yes, with care. Microsoft runs an EU Data Boundary and Foundry offers EU Data Zone deployments pinned to European regions such as Sweden Central and Germany West Central, which keep storage and processing inside the geography. Global Standard deployments can route prompts outside the chosen region, so pick Data Zone or Regional deployments for personal data, sign the Microsoft Products and Services Data Protection Addendum (DPA), pick a lawful basis under GDPR Art. 6, keep a record of processing under Art. 30, run a DPIA where risk is high under Art. 35 and train anyone building AI steps under EU AI Act Art. 4. Superkind is hosted and processed in Europe, GDPR compliant, aligned with the EU AI Act, never trains models on your data and provides the DPA through the Trust Center at trust.superkind.ai.

Superkind has public prices, as of October 2026. Free costs EUR 0 with a one-off up to EUR 150 in credits, no credit card. Business starts at EUR 50 per month for 25,000 credits (EUR 100 for 50,000 credits is the popular step, then EUR 200 and EUR 500) for the whole company with unlimited users, so no price per seat; unused credits carry over one month and you can cancel monthly. A question uses 10-50 credits, a task 50-500 and a run of your own agent 1,000-5,000, so EUR 100 covers up to 1,000 tasks or 5,000 questions. Enterprise has a custom price and all prices are excl. VAT. Superkind credits are not the same as Azure tokens or PTUs.

Four big ones: Foundry Tools (file-search vector storage per GB per day with the first 1 GB free, code interpreter per session, web search per 1,000 transactions, custom search per 1,000 transactions); Azure AI Search for Foundry IQ retrieval, billed as its own Azure service; Azure storage and Azure Machine Learning if your agent uses them; and implementation (partner day rates to build agents, connectors and knowledge bases, plus the owner time to curate the sources an agent is grounded in). Add fine-tuning (training hours plus hosting per hour per deployed model) if you fine-tune. Always ask Microsoft for a quote that lists Foundry Tools, Azure AI Search and partner work next to the model tokens.

Yes, for example with Superkind. Your team writes to Superkind in Outlook, Microsoft Teams, Slack or the web app, and it does the task in your programs through 1,500+ integrations such as SharePoint, Excel, Outlook, HubSpot, Personio, DATEV and lexoffice, with AI models from OpenAI, Anthropic, Mistral and Google, all processed in Europe. Superkind asks for approval on the steps you choose. The Business plan covers the whole company with unlimited users from EUR 50 per month, and you can start for free at app.superkind.ai without an Azure subscription or any DevOps setup.

Henri Jung, Co-founder at Superkind
Henri Jung

Co-founder of Superkind, the AI employee teams write to in Outlook, Teams and Slack. Henri helps companies, from small teams to the Mittelstand, hand their routine work to AI - without new tools, consultants or an in-house AI team.

Send Superkind your first platform task.

Start for free in Outlook, Teams or Slack, without an Azure subscription or a Foundry deployment. Or see in 30 minutes how it runs on your estate.