Azure AI Foundry Pricing: What Startup Credits Cover — and What They Don’t

What the invoices actually showed

Microsoft for Startups gave me credits that let me build an AI stack I could not otherwise have afforded. That part is real. I would recommend the program to another founder tomorrow.

Azure AI pricing is not a single rate card. It depends on the model and on who bills for it.

Then two invoices arrived.

The credits had worked on everything else. They had not worked on Claude.

I am not writing this to complain. I am writing it because the warning exists, it is just in the wrong place — and moving one sentence would have stopped me deploying.

The credits were working

Here is the part most people miss. It was not that my credits ran out.

The credits were still being applied in September — ₹87,574.52 of them. Every first-party line on the invoice (AI and Machine Learning, Storage, Web, Compute, Networking, Containers) was charged and then credited back.

The only line that survived was SaaS: the Claude Marketplace subscription.

So this was not a balance problem. It was a coverage problem. The credits worked exactly as designed on everything except the one product line I did not know was excluded.

Why this happens: first-party versus Marketplace

Microsoft sells some models directly. Azure OpenAI GPT models are the obvious example. Azure OpenAI costs are billed as standard Azure consumption, and startup credits apply.

Other models in the same catalog — including every Claude model — are published by partners and billed through Microsoft Marketplace. They are deployable from Foundry, they look like Foundry models, and they appear on your Azure bill. But the billing source is different, so the credits do not apply.

Anthropic is the first name in its “not covered” list.

The same guidance appears in “Best ways to use your startup credits”: “Models from third-party providers (including Anthropic) and partners via Microsoft Marketplace aren’t covered.”

So the rule is not secret. It is published, it is specific, and it names the provider.

That is exactly why the placement matters.

The warning exists. It is six scrolls down.

When I checked the Foundry model page in October 2026, the top of the page looked like this:

There was no “credit eligibility” field. There was no “Marketplace-billed” label. The word “Azure” appeared once, as “Hosted on Azure.”

The Deploy button sat at the top right, above all of it.

The warning was on the page. It was inside a collapsed Pricing accordion, several screens down, below “Use cases”:

That is the right sentence. It is in the wrong place.

A founder deploying a model reads the panel next to the button. They do not open a collapsed pricing section below the fold before clicking Deploy — especially when the panel already tells them the model is hosted on Azure, and especially when the model next to it (Azure OpenAI) is covered by credits.

Why “Hosted on Azure” is the problem

The Quick facts block is the only summary a founder sees before deploying. It already contains two commercial statements:

• Hosted on Azure
• Pricing → View pricing

So the block is not purely technical metadata. It already tells you where the model runs and links you to what it costs. It simply does not tell you who bills you — or whether your credits apply.

If the panel can say where the model is hosted, it can say who bills for it.

That is the entire gap. It is one row.

This is not a one-off

On Microsoft’s own community platform, the same invoice appears month after month. The largest reported cases:

Ten cases, ten months, all on Microsoft’s own forum. The full dated table — with the quotes, the escalation loop, and what each founder did — is in Unexpected Claude Charges on Azure: Why They Happen and How to Fix Them.

The part that is easy to miss: for a funded startup, a $17,000 invoice is painful. For a bootstrapped founder, it can be the runway.

A billing rule that only becomes visible after the invoice is not a pricing decision. It is a trapdoor.

Before you deploy — or after you’ve been billed

There are two practical guides, and they live on one page:

• The five-point pre-deployment check — how to confirm a model is Marketplace-billed before you click Deploy.
• The rescue playbook — what to do if the invoice has already arrived, who to contact, and what each party will tell you.

Read the full checklist and escalation guide →

The short version: open the model card, check whether it is in “Direct from Azure” or “Models from Partners and Community,” and read the Pricing section for the sentence about Azure credits. If your credit balance is not moving while you generate tokens, that is the warning.

Five changes we would suggest to Microsoft

None of these is a policy change. All of them are placement changes.

Any one of these would have stopped me before the first deployment.

All five are a day of front-end work.

What we are not saying

We are not saying Microsoft hid the policy. It is published, and it names Anthropic explicitly.

We are not saying anyone acted in bad faith. The support engineers we dealt with were professional, and the policy they applied is written down.

The program’s intent is good. The credits are real, and they let founders build things they could not otherwise afford. But goodwill is fragile. When a founder who was supposed to benefit ends up with a bill they cannot pay, part of that goodwill is lost — not because the policy is wrong, but because the warning arrived after the decision. That is the real cost of a disclosure gap.

We are saying the disclosure is in the wrong place. Microsoft’s own statement to The Register was that customers should “rely on official documentation.” Founders do. The problem is that the deploy page does not tell them there is a document to read.

Right of reply

Microsoft and Anthropic are invited to respond. Any substantive response will be published here in full, with a date.

Corrections are welcome and will be dated.

FAQ

Related

• Unexpected Claude Charges on Azure: Why They Happen and How to Fix Them — the full rescue playbook and the ten-case table.
• Microsoft for Startups: A Founder’s Field Guide — what the program gives you, and what to check before you deploy.
• Best Small Business Tools Stack Under $100 in 2026 — how to keep the rest of the stack inside a budget while you absorb a Marketplace charge.
• Best AI Content Stack for a New Business in 2026 — how to choose AI tools when the billing model matters as much as the output.

Arvind Jadli is the founder of Sovor and GSA Growth. He is a Microsoft for Startups participant. This article is based on his own invoices, his support correspondence, and publicly linked founder reports. He has not been paid by Microsoft or Anthropic.