Most AI Foundry pricing analysis online is one model at a time. "Is GPT-5 cheaper on Foundry?" Yes, same price. "Is DeepSeek?" Different question, different answer. The interesting picture only shows up when you put the whole catalogue on one chart.
Here's the full markup table for May 2026. Foundry rates pulled from actual Azure usage CSVs on a pay-as-you-go subscription, direct-API rates from cloudprice.net and the providers' own pricing pages.
The full table
| Model | Foundry In $/M | Foundry Out $/M | Direct API Out $/M | Azure markup |
|---|---|---|---|---|
| gpt-5-pro | $15.00 | $120.00 | $120.00 | 1× (no markup) |
| gpt-5 / gpt-5-codex | $1.25 | $10.00 | $10.00 | 1× |
| gpt-5-2 / 5-2-codex | $1.75 | $14.00 | $14.00 | 1× |
| gpt-4o | $2.50 | $10.00 | $10.00 | 1× |
| gpt-4.1 | $2.00 | $8.00 | $8.00 | 1× |
| o4-mini | $1.10 | $4.40 | $4.00 | 1× |
| claude-opus-4-6 | $5.00 | $25.00 | $25.00 | 1× |
| claude-sonnet-4-6 | $3.00 | $15.00 | $15.00 | 1× |
| mistral-large-3 | $0.50 | $1.50 | $1.20 | 1.25× |
| llama-4-maverick-fp8 | $0.25 | $1.00 | $0.85 | ~1.2× |
| llama-3.3-70b | $0.71 | $0.71 | $0.20 | 3.5× |
| deepseek-v3-2-sp | $0.58 | $1.68 | $0.50 | 3× (promo) |
| deepseek-v3-1 | $1.23 | $4.94 | $0.50 | 10× |
| deepseek-r1 | $1.35 | $5.40 | $0.40 | 13× |
| deepseek-v4-flash | $1.03 | $4.12 | $0.28 | 15× (GA May 2026) |
Source: Azure usage CSV (pay-as-you-go subscription, May 2026), cloudprice.net, api-docs.deepseek.com.
The pattern
There are three groups, and they tell three different stories.
Group 1, OpenAI and Anthropic: no markup. Whatever you pay direct, you pay on Foundry. This is genuinely surprising the first time you see it. Microsoft doesn't put a margin on top of the Azure OpenAI rate cards. Anthropic Claude on Foundry, even though inference physically routes to Anthropic infrastructure, comes in at the same per-token rate as Anthropic's own API. The markup model isn't extracted from the per-token bill; it's elsewhere (compute commit, support contract, Marketplace billing). For procurement teams this is the easy part: zero pricing penalty for staying inside the Microsoft DPA, at least for the OpenAI/Anthropic surface.
Group 2, Mistral and Llama: light markup, mostly explainable. Mistral Large 3 at 1.25× and Llama 4 Maverick at ~1.2× over direct partner pricing is operational overhead. Microsoft hosting infrastructure, Marketplace contracts, support escalation paths. Llama 3.3 70B sits awkwardly at 3.5× because the direct-API price floor for that model is so low ($0.20/M out from some providers) that any operational margin looks proportionally large.
Group 3, DeepSeek: the sovereignty tax. This is where the markup gets real. DeepSeek V4 Flash at 15× the direct API rate. R1 at 13×. V3.1 at 10×. And every one of these is a model that's nominally Class A on Foundry: Direct from Azure, Microsoft is data processor, runs on Azure infrastructure.
That 15× is what you're paying for. It's not Microsoft skimming arbitrarily. It's the cost of moving an open-weights Chinese-published model onto Azure infrastructure with US/EU data-processor commitments, Microsoft's compliance posture, and Microsoft's Anthropic-style sub-processor disclosures. DeepSeek's direct API has none of those. It's hosted in China under Chinese data laws. The two aren't the same product despite using the same weights.
What you're actually paying for at 15×
Walk through what changes when you pick deepseek-v4-flash on Foundry vs deepseek.com:
| Dimension | DeepSeek direct API | DeepSeek on Azure Foundry |
|---|---|---|
| Inference location | China (mainland) | Azure region you select |
| Data processor under your DPA | DeepSeek (Chinese) | Microsoft (DPA terms you already have) |
| Authentication | API key | Entra ID + Managed Identity, key-less option |
| Network controls | Public internet only | Private Endpoint, VNet integration |
| Logging | Provider-side | Azure Monitor / Log Analytics native |
| Compliance certifications | Provider's own | Inherits Azure SOC2/ISO/HIPAA scope |
| SLA | Provider's | Azure-backed |
| Billing | Separate vendor invoice | Single Azure billing line, MCA-compatible |
| Procurement | New vendor onboarding | Existing Microsoft contract |
| Data sovereignty story to your auditor | Hard | Easy |
For a UK financial services client, that 15× is the answer to a question they'd otherwise have to refuse to engage with at all. A Chinese inference endpoint isn't on the menu for regulated workloads regardless of price. A Microsoft-fronted deployment of the same model is.
For a personal side project? Pay direct.
When the sovereignty tax is the wrong choice
The 15× makes Foundry-hosted DeepSeek the wrong tier in three specific situations:
-
You're already on a different sovereignty story. If your stack runs on AWS Bedrock for the rest of your AI workload, switching to Foundry just for DeepSeek doubles your governance surface. Stay on Bedrock if it carries the model, or run direct.
-
Your workload is heavy enough that the markup dominates the bill. A 100M-token-per-month DeepSeek V4 Flash workload costs $412 on Foundry vs $28 direct. At that scale the engineering effort to bring inference in-house (managed compute on Foundry, or self-hosted spot GPU) starts paying back fast.
-
You'd actually pick Mistral if you priced it side-by-side. Mistral Large 3 on Foundry at $1.50/M out is 2.7× cheaper than V4 Flash at $4.12, and scores 86% to V4's 90% on independent IaC benchmarks. Most teams that default to V4 Flash for cost reasons haven't compared the alternatives. Output-token-dominated bills (Foundry's typical workload) make the Mistral path the cleaner default.
The macro pattern
There's a simple rule emerging from this table: the more reputational risk a publisher carries for an enterprise procurement team, the bigger the Foundry markup is. OpenAI and Anthropic: zero markup, because Microsoft already has those companies fully integrated into the procurement story. DeepSeek: 15×, because Microsoft is doing real work to make a Chinese-published model deployable in a Microsoft-shop tenant. Mistral and Llama in between, where the publisher is European/American but Microsoft still has to operate the infrastructure.
That's not a value judgement about the models. It's a description of what enterprise procurement is willing to pay to make a model "boring" in the same way that Azure Storage is boring. For most enterprise clients, "boring" is the goal, and 15× of DeepSeek's $0.28/M is worth it.
For everyone else, the direct API is right there.
What this means for FinOps
If you're advising on AI spend across a portfolio of Foundry tenants, the markup pattern reshapes the model-selection conversation:
- Don't compare Foundry vs Foundry on cost alone for partner models. Pick the right model first, then check whether direct-API would make sense for that specific workload.
- Tag every deployment with its hosting class (Class A direct, Class B partner serverless, Class C marketplace). Costs and residency have different shapes per class.
- The output-token rate is the only number that matters for code generation and content workloads. Input is a rounding error.
- DeepSeek and Llama 3.3 70B get FinOps scrutiny as the high-markup partners. OpenAI and Anthropic don't need it.
Want this analysis applied to your actual Azure usage? Request a free FinOps assessment and we'll pull your real markup numbers from Cost Management.