Microsoft Ignite 2026: What Does an Enterprise AI Agent Cost? Azure Resources, Model Calls, and Runtime

The monthly cost of an enterprise AI agent comes down to three choices: which model tier is the default, whether runtime is billed only while a session is active, and which steps never run on Azure. Microsoft Ignite 2026 is in San Francisco from November 17 to 20, with an online option. As of October 10, 2026, the opening keynote has not happened. This article does not treat unpublished prices as fact. It uses rates Microsoft has already posted, plus one worked example whose assumptions are written down.

Sources: the Microsoft Security guide to Ignite 2026 (September 30), GPT-5.6 pricing in Foundry, the Foundry Agent Service pricing page, how hosted agents are billed, Copilot Studio credit packs, and per-minute Azure Pipelines agents (October 1). The monthly method is the same as costing AI programming by the task. For OpenAI’s sandbox bill, see what server an agent needs after DevDay.

Prices that already apply

Microsoft’s own security blog already sets the Ignite security thread. Security Pre-Day is November 16, 1:00–5:00 p.m. PT. On November 17, 10:00 a.m.–12:00 p.m., Satya Nadella and Judson Althoff give the opening keynote at Chase Center. The catalog already includes Agent 365, finding and protecting local agents, and access control for non-human identities. Those sessions change governance and permissions. They do not by themselves change the token rate.

Model prices landed before the conference. The Foundry post lists Standard Global, short-context rates for the three GPT-5.6 tiers, in US dollars per million tokens:

The promotion runs from September 1, 2026, through at least November 30: input drops to $4.00 and output to $20.00. The note does not change cached input or cache writes, so the promo total below still uses cached input at $0.50. Data Zone and Priority Processing are “ask your sales rep.” They are not columns in the public table. The Asia-Pacific Data Zone is already generally available. A region option is not a discount.

One more product uses the word “agent” and does not belong on the same purchase order. Since October 1, GitHub-hosted agents in Azure Pipelines bill per minute. macOS Standard is $0.062/minute, macOS XLarge is $0.102/minute, and Linux 8-core is $0.022/minute. That is a CI machine, not the reasoning loop inside Foundry.

Four lines on one bill

Split “what does it cost to deploy an agent?” into four lines. The split still holds after the keynote:

The Foundry pricing page says there is no extra charge to create or run Foundry-native agents with prompts and workflows. Model tokens, tools, and knowledge connections are separate. Hosted agents are for external frameworks such as Microsoft Agent Framework, LangGraph, and OpenClaw, billed as container CPU and memory. The docs add a capacity point that is easy to miss: the CPU and memory on an agent version describe one session, not the agent’s total footprint. Billing sums every active session, so a larger per-session size multiplies cost by concurrency.

Model calls: pick the tier first

This worksheet is an order of magnitude, not a quote. Assume an internal knowledge agent:

• 22 workdays, 500 conversations a day, 11,000 a month.
• Each turn: 6,000 fresh input tokens, 2,000 cached input tokens, 1,200 output tokens.
• Cache hits use the cached-input rate, not the cache-write rate. Short-context list prices. Do not apply them to a million-token prompt.
• No code interpreter and no Bing. Both are separate meters. Use the calculator for those dollars.

Monthly tokens: 66 million fresh input, 22 million cached input, 13.2 million output.

On this mix, Terra is exactly 10× Luna and list-price Sol is exactly 25× Luna, because input, cached input, and output all scale by the same factor. The promotion moves Sol from $737 to $539. It does not apply one more neat discount to the cache. When traffic grows, change the default model before you debate another container.

These figures cover a question-and-answer turn. Once the agent stuffs tool definitions, multi-step reasoning, and a long document into the same call, input moves from 6,000 tokens to tens of thousands. Toolboxes in Foundry are generally available so a request does not carry the entire tool catalog. Cutting that list usually shows up on the bill before another GPU does. For buy-versus-rent on the GPU itself, see buying a card or renting one.

Hosted runtime: sessions, not idle machines

For a native agent the runtime line is $0, and the model table is the whole story. Add a hosted container when you need your own runtime, an isolated sandbox, or an external framework. Microsoft’s docs confirm billing on CPU and memory during active sessions, with scale-to-zero when idle. The dollar rates used here are the April 2026 citations of $0.0994 per vCPU-hour and $0.0118 per GiB-hour. Common sandbox sizes fall between 0.25 and 2 vCPU and 0.5 to 4 GiB. Replace the pair in the Azure pricing calculator before you order.

Keep the 11,000 conversations and assume each one is active for 2 minutes:

• Session-hours = 11,000 × 2 ÷ 60 = 366.67.
• 1 vCPU: 366.67 × $0.0994 = $36.45.
• 2 GiB: 366.67 × 2 × $0.0118 = $8.65.
• Total $45.10. Add the model: about $29.48 + $45.10 = $74.58 on Luna, and about $737 + $45.10 = $782.10 on list-price Sol.

Raise the session to 2 vCPU and 4 GiB and both lines double. Runtime becomes $90.20. Twenty-way concurrency does not multiply 366.67 hours by 20. Total vCPU-hours are the sum of session durations. What multiplies is the size you set on the version. Peak concurrency decides how many sandboxes exist at once. It does not change the hour total of a 2-minute bill.

The other mistake is to buy 2 vCPU and 4 GiB as if they run all 730 hours, with no scale-to-zero: 730 × 2 × $0.0994 = $145.12, and 730 × 4 × $0.0118 = $34.46, together $179.58. You pay that with zero conversations. Against 366.67 active hours a month, always-on is about 4× the per-session bill. At lower traffic the multiple is larger.

Copilot Studio credits are not tokens

Agents owned by a business team often skip the Foundry token table and land in Copilot Studio. Microsoft’s pack is 25,000 Copilot Credits for $200 a month, which is $0.008 per credit. An Azure subscription is required. Prepaid commit units can save up to 20 percent. That is Microsoft’s statement about the pre-purchase tier. A pay-as-you-go meter exists, but the public page does not print a unit price equal to the pack. After prepaid credits run out, use the subscription invoice.

Credits per reply depend on task complexity. Microsoft does not publish a fixed “one message = N credits,” so $29.48 cannot prove that Studio is cheaper or more expensive. Pick a credit count you still have to measure, then set it next to the Foundry model line:

• If a one-week sample averages 2 credits: 11,000 × 2 = 22,000 credits, inside one pack, $200 a month, with 3,000 credits left. That is above Luna at $29.48 and below Terra at $294.80.
• If the average is 8 credits: 88,000 credits, or 3.52 packs. At the pack unit price that is 88,000 × $0.008 = $704, close to list-price Sol at $737. Buying whole packs means 4 × $200 = $800. Filling the remainder on the meter depends on the invoice.

Both 2 and 8 are assumptions, there to show scale. They are not a product quota. Before go-live, sample a week and record credits per reply, whether an action ran, and whether the channel was external. Without that week, an annual budget is a guess about the meter.

The Azure resources beside the model

After the model and the container, production usually has more lines. The pricing page names them. The dollars sit in the calculator, and this article does not invent a figure Microsoft has not posted:

• Foundry IQ, Azure AI Search, SharePoint, and Fabric data agents: the knowledge layer, often with its own retrieval and storage charges.
• File-search vector storage (the pricing page includes 1 GB free, then a per-GB-day rate), code interpreter per session, and web or custom search per 1,000 transactions.
• Logic Apps connectors, private endpoints, a virtual network, Key Vault, and a log workspace. Before you ship every tool-call trace into Log Analytics, look at one week of ingestion. The log line can exceed Luna’s $29.48.

On identity, the Ignite catalog already puts non-human identities and Agent 365 on the security track. A separate Entra identity for an agent usually comes from the Microsoft agreement you already have. It is not a row on the GPT-5.6 token table. Design the permissions before the conference. Keep the cost on the four lines above.

macOS work is a separate bill

A Foundry hosted sandbox runs the agent loop inside a Linux container. It does not compile an Xcode project, notarize a build, or drive a device. Those steps stay on Apple Silicon, and they get their own budget line. See splitting Mac development from GPU capacity and the Mac mini agent checklist.

Two published Mac prices should not be added to Foundry’s $45.10 and then called an “agent server”:

• Per-minute CI: 8 builds a workday, 15 minutes each, 22 days, 2,640 minutes. macOS Standard at $0.062 × 2,640 = $163.68. The same minutes on XLarge are $0.102 × 2,640 = $269.28. Linux 8-core is $0.022 × 2,640 = $58.08. A pipeline with no signing and no device should not sit on the macOS SKU.
• A whole machine by the month: a Zilmac cloud Mac mini M4 with 16 GB starts at $96.1/month (the plan page is the live price). One machine fits serial builds and desktop acceptance. Parallel pipelines stay on the per-minute meter, or on another node.

A Python or Node agent can stay on Foundry or on your own Linux. Renting a cloud Mac to host that loop spends a monthly machine on work that ends with the session. Putting notarization into a hosted Linux container saves the $45 and still does not produce a build you can ship.

Numbers to record before November 17

The keynote can rename the default model the same day. The budget table can have its shape before that:

1. Export one week of tokens, split by model, input, cached input, and output. Replace 11,000 with your count and rebuild the $29.48 table.
2. Label every agent: Foundry-native, hosted container, Copilot Studio, or a Mac step. The label decides which line appears.
3. For hosted containers, write down vCPU, GiB, and real active minutes per session. Check first for a size that never scales to zero.
4. Sample Copilot Studio for a week and compute credits per reply. Without that number, do not turn a $200 pack into “dollars per person per month.”
5. At the conference, ask for Data Zone, Priority Processing, per-transaction tool prices, and how to export credits from a trace. Those figures are not on the public table today.
6. Promo Sol at $4 / $20 runs through at least November 30. Use the promo rate on the November invoice. December returns to $5 / $30 unless the docs move the date.

The day after the keynote, swap the model ID and the unit price. The four-line structure stays. Whether to buy a GPU still depends on whether the hours pay the card back, not on a conference slogan.

FAQ

Can the default model cost more than the container?

On this worksheet of 11,000 conversations a month, about 2 minutes each, yes. Luna’s model line is $29.48 and a 1 vCPU + 2 GiB hosted runtime is $45.10. They are close. At Sol list price the model is $737 and the runtime is still $45.10. The runtime line grows when you enlarge the sandbox or leave it on for 730 hours. A wrong default model moves the bill first.

Should purchasing wait for Ignite?

Do not hold a native agent that is ready to ship just to wait for a unit price that has not been announced. Before November 17, Standard Global short-context rates, the $200 credit pack, and the Pipelines per-minute prices are already in effect. Ignite is the place to bring one week of traces and ask about Data Zone and tool prices. A 12-month always-on GPU or container can wait until your own session-hours are on paper.

How many credits does one chat use?

Microsoft does not publish a fixed deduction. A pack is 25,000 credits for $200, or $0.008 per credit. Credits per reply vary with complexity. After a week of real traffic gives you the average, you can set it next to Foundry’s range from $29.48 to $737. The 2-credit and 8-credit cases in this article only show scale.

When do you need a cloud Mac?

When the work includes Xcode, notarization, a device, or a bug that only reproduces on macOS. Questions and ordinary scripts stay on Foundry’s model and hosted runtime. If CI is 8 builds a day at 15 minutes, macOS Standard is about $163.68 a month. For serial work that also needs a desktop, compare a whole machine from $96.1 a month. Do not add those two bills into one “Azure agent” account.