Something interesting is happening inside Microsoft’s AI strategy right now, and it’s more relevant to your bottom line than it might first appear.
In January 2025, Microsoft added DeepSeek R1 to the Azure AI Foundry model catalogue, making the Chinese open-weight model available to developers and enterprises through Microsoft’s cloud platform.
Then, in mid-June this year, Axios reported that Microsoft is now evaluating DeepSeek V4 as a lower-cost model option for Copilot Cowork, its new agentic productivity tool that went generally available on 16 June 2026.
Microsoft told Axios it has already fine-tuned a version of the model and expects to confirm its decision in the coming weeks. Both moves point to the same thing: Microsoft is no longer building Copilot around a single AI model, and the reason is almost entirely about cost.
That shift matters to any business leader thinking about where AI fits in their budget. But it also raises some questions worth asking, especially if your business operates in a regulated industry.
Why Microsoft is Mixing Models
To understand the strategy, you need to know how AI pricing actually works.
Five-step infographic showing how AI models process tokens and how token-based billing is calculated.
When you send a request to an AI model, asking it to summarise a document, draft an email, or answer a complex question, the model processes it using what are called tokens, essentially small chunks of text. You’re billed per million tokens processed.
That sounds abstract, but the numbers are striking. Anthropic’s Fable 5, its most capable publicly available model, costs around USD 50 per million output tokens on the API. DeepSeek V4 Pro costs around USD 0.87 per million output tokens, a price gap of roughly 57 times between the two.
For a single user sending occasional prompts, that gap is irrelevant. But Copilot Cowork isn’t designed for occasional prompts. It’s built for autonomous, multi-step tasks: agents that research, draft, review, update records and follow up, often dozens of times a day, across an entire business.
At that volume, model costs add up fast. Microsoft has been blunt about this. The company told Axios that Copilot Cowork “can’t work on unlimited pricing” because some users are running hundreds of tasks per week.
Adding a cheaper model for routine tasks isn’t a quality compromise. It’s how you make the economics work.
What DeepSeek Is and Where It Comes From
DeepSeek is a Chinese AI company that released a series of open-weight models, meaning the underlying model weights are publicly available for anyone to use, modify or deploy.
That’s different from the closed, proprietary models from OpenAI and Anthropic, where the inner workings stay private. DeepSeek’s R1 and V4 models have performed well on a range of technical benchmarks and attracted serious attention in the AI community for their efficiency relative to their cost.
That combination of performance and low price is exactly what Microsoft is interested in.
The concern many business leaders raise immediately is an understandable one: should I be worried about a Chinese-developed model handling my business data?
Microsoft’s answer is that any DeepSeek option inside Copilot Cowork would be a fine-tuned version hosted entirely on Azure, inside Microsoft’s cloud infrastructure. Your data stays within your existing Microsoft 365 security boundary and compliance controls. It doesn’t go to DeepSeek’s servers, and it doesn’t leave your Microsoft tenant.
That’s the same principle that applies to every model Microsoft surfaces through Azure AI Foundry, whether it’s OpenAI’s GPT-5.5, Anthropic’s Claude, or an open-source model. Microsoft controls the hosting, the security layer, and the data handling. The model is the intelligence engine, not the data keeper.
For most businesses, that reassurance is enough. But it isn’t the full story for everyone.
What About Regulated and Defence-Adjacent Businesses?
If your business operates under a compliance framework like DISP (Defence Industry Security Program) or requires an IRAP (Information Security Registered Assessors Program) assessment, the conversation goes further than where your data is hosted.
These frameworks are designed for businesses that handle sensitive government, defence, or critical infrastructure information.
Under DISP, businesses engaging in defence tenders are assessed across personnel, physical, cyber, and supply chain security domains, and for protected, secret, or higher-classified workloads, public cloud AI tools generally aren’t on the table at all.
IRAP assessments evaluate cloud services against the Australian Government Information Security Manual (ISM), and Microsoft completed updated IRAP assessments for Azure, Dynamics 365 and Microsoft 365 at the “Protected” level in early 2026, supporting Australian government workloads. That’s a meaningful credential for those environments.
But here’s what matters: IRAP certification covers the Microsoft platform, not every model that runs on it. A Chinese-developed model running inside an IRAP-assessed Azure tenancy doesn’t automatically inherit that certification for your workload.
The provenance of the model weights, the training data used, and the potential for embedded vulnerabilities are separate questions that your risk and compliance team would need to work through.
Independent security evaluations add weight to those concerns. In September 2025, the Center for AI Standards and Innovation at NIST evaluated several DeepSeek models and found significant shortcomings.
Agents built on DeepSeek’s R1 model were, on average, 12 times more likely than evaluated US frontier models to follow malicious instructions in agent hijacking scenarios. The same evaluation found that DeepSeek’s most tested model responded to 94% of overtly malicious requests when a common jailbreaking technique was used, compared with 8% for US reference models.
That’s a meaningful gap for any environment where agents operate with real permissions across business systems.
What this means in practice
None of this means DeepSeek-on-Azure is automatically unsuitable for regulated businesses. It means the decision requires more than a hosting check. A few questions worth asking before any model option goes near sensitive workloads:
Does your compliance framework require you to assess the provenance of AI model weights, not just the hosting environment?
Have you mapped which Copilot workloads will run on which models, and do those mappings align with your data classification policies?
If Microsoft introduces a new model option that is “off by default,” do you have a process to review and explicitly approve or block it before it reaches your users?
The good news is that Microsoft’s multi-model platform is designed to give businesses that choice.
You don’t have to use DeepSeek if it doesn’t suit your risk profile. GPT-5.5, Anthropic’s Claude, and Microsoft’s own MAI models are all available within the same Copilot environment, and Microsoft’s admin tooling lets you control which models are accessible to your users.
This Is Part of a Bigger Shift
The DeepSeek story is really just one thread in a larger move Microsoft has been making through 2026. As I wrote when covering Microsoft Build in June, the company is explicitly building a multi-model future where different models serve different jobs.
Microsoft’s official position, published on 16 June, is that both Microsoft 365 Copilot and GitHub Copilot are “model-diverse by design”, meaning the platform won’t lock you into a single provider. GPT-5.5 Instant, Claude Opus 4.8 and other models already run inside Copilot Chat and Copilot Studio today.
The addition of a lower-cost open-source option for agent tasks follows that same logic: match the model to the job and the cost to the value.
This is actually how mature technology platforms tend to work. Your business probably already uses different cloud services for different workloads.
You don’t host everything on the most expensive option when a cheaper one does the job just as well. AI is heading in the same direction, though with more compliance dimensions to manage.
What It Means for Your AI Costs
If you’re running Microsoft 365 Copilot now, or planning to adopt Copilot Cowork as it scales, here’s what this shift means in practical terms.
Routine tasks will likely get cheaper to run
The tasks that Copilot Cowork handles most often, like summarising documents, drafting standard communications, and pulling data from Microsoft 365, don’t always need the most powerful model available.
Routing those to a lower-cost model while reserving frontier models for complex reasoning, legal review, or strategic analysis is a sensible design. Microsoft is building that routing logic into the platform.
Your data protection doesn’t change
When Microsoft runs any model on Azure, your business data stays inside the same Microsoft 365 security perimeter it’s always been in. Permissions, sensitivity labels, and compliance policies still apply.
The model processes the request; it doesn’t store your data, and it doesn’t share it externally. I covered the governance layer in detail when we looked at securing AI agents with Entra Agent ID and Purview, and that framework applies here too.
Cost visibility is becoming non-negotiable
One thing the DeepSeek-in-Copilot story makes clear is that AI costs are no longer just a subscription line item. As agents run more tasks autonomously, usage-based charges will show up more frequently.
Microsoft introduced a Cost Management Dashboard in the Microsoft 365 Admin Centre as part of the Work IQ API general availability, and it’s worth getting familiar with that tooling before your agent usage scales.
How to Think About Open-Source Models in Your AI Strategy
The conversation around open-source AI models is often framed as “open source versus Microsoft”, but that’s not really how it plays out in practice for most businesses. At CG TECH, I talk about this a lot with clients who are designing an AI operating model that works across multiple tools and providers.
Open-source models like DeepSeek aren’t an alternative to Microsoft. They’re one option within a broader portfolio that Microsoft itself is curating.
The more useful question isn’t “should we use open-source models?” It’s “which tasks are worth paying frontier model prices for, and which ones aren’t?”
Some good places to start asking that question:
High-volume, lower-complexity tasks like scheduling, data extraction, standard document formatting, and routing queries are strong candidates for cost-efficient models
Complex reasoning, legal review, sensitive decisions, or nuanced client communications are scenarios where frontier model quality is worth paying for
Any task where an error has real consequences: regardless of model, a human should still be reviewing outputs before they’re acted on
Any workload touching classified, defence, or sensitive government data needs a compliance review before you make any model decision
The businesses I see getting the most from AI right now aren’t the ones using the most expensive tools. They’re the ones being deliberate about matching the right model to the right task, and knowing which tasks require a higher bar.
Where Things Are Heading
Microsoft hasn’t confirmed a release date for the DeepSeek V4 option in Copilot Cowork; the June reporting describes it as being tested, with availability expected in the “coming weeks.” It’s worth watching, because the pricing implications for businesses running agents at scale could be meaningful.
What’s already confirmed is the direction of travel: Microsoft 365 Copilot is a multi-model platform, model costs vary significantly across providers, and Microsoft is building the tooling to help businesses govern and manage that spend. The DeepSeek integration, if it lands, will be the most visible example yet of how far that strategy has come, and it’ll be optional, which matters.
For businesses operating in regulated industries, the message isn’t “avoid DeepSeek.” It’s “make the decision deliberately, with your compliance obligations in front of you.” That’s a different conversation to the one most AI vendors want to have.
If you’d like to talk through how your Microsoft 365 environment is set up for the shift to usage-based agent pricing, or what a practical multi-model strategy looks like for your business, reach out to the CG TECH team. It’s a conversation worth having before the bill, or the audit, surprises you.
About the Author
Carlos Garcia is the Founder and Managing Director of CG TECH, where he leads enterprise digital transformation projects across Australia.
With deep experience in business process automation, Microsoft 365, and AI-powered workplace solutions, Carlos has helped businesses in government, healthcare, and enterprise sectors streamline workflows and improve efficiency.
Something interesting is happening inside Microsoft’s AI strategy right now, and it’s more relevant to your bottom line than it might first appear.
In January 2025, Microsoft added DeepSeek R1 to the Azure AI Foundry model catalogue, making the Chinese open-weight model available to developers and enterprises through Microsoft’s cloud platform.
Then, in mid-June this year, Axios reported that Microsoft is now evaluating DeepSeek V4 as a lower-cost model option for Copilot Cowork, its new agentic productivity tool that went generally available on 16 June 2026.
Microsoft told Axios it has already fine-tuned a version of the model and expects to confirm its decision in the coming weeks. Both moves point to the same thing: Microsoft is no longer building Copilot around a single AI model, and the reason is almost entirely about cost.
That shift matters to any business leader thinking about where AI fits in their budget. But it also raises some questions worth asking, especially if your business operates in a regulated industry.
Why Microsoft is Mixing Models
To understand the strategy, you need to know how AI pricing actually works.
When you send a request to an AI model, asking it to summarise a document, draft an email, or answer a complex question, the model processes it using what are called tokens, essentially small chunks of text. You’re billed per million tokens processed.
That sounds abstract, but the numbers are striking. Anthropic’s Fable 5, its most capable publicly available model, costs around USD 50 per million output tokens on the API. DeepSeek V4 Pro costs around USD 0.87 per million output tokens, a price gap of roughly 57 times between the two.
For a single user sending occasional prompts, that gap is irrelevant. But Copilot Cowork isn’t designed for occasional prompts. It’s built for autonomous, multi-step tasks: agents that research, draft, review, update records and follow up, often dozens of times a day, across an entire business.
At that volume, model costs add up fast. Microsoft has been blunt about this. The company told Axios that Copilot Cowork “can’t work on unlimited pricing” because some users are running hundreds of tasks per week.
Adding a cheaper model for routine tasks isn’t a quality compromise. It’s how you make the economics work.
What DeepSeek Is and Where It Comes From
DeepSeek is a Chinese AI company that released a series of open-weight models, meaning the underlying model weights are publicly available for anyone to use, modify or deploy.
That’s different from the closed, proprietary models from OpenAI and Anthropic, where the inner workings stay private. DeepSeek’s R1 and V4 models have performed well on a range of technical benchmarks and attracted serious attention in the AI community for their efficiency relative to their cost.
That combination of performance and low price is exactly what Microsoft is interested in.
The concern many business leaders raise immediately is an understandable one: should I be worried about a Chinese-developed model handling my business data?
Microsoft’s answer is that any DeepSeek option inside Copilot Cowork would be a fine-tuned version hosted entirely on Azure, inside Microsoft’s cloud infrastructure. Your data stays within your existing Microsoft 365 security boundary and compliance controls. It doesn’t go to DeepSeek’s servers, and it doesn’t leave your Microsoft tenant.
That’s the same principle that applies to every model Microsoft surfaces through Azure AI Foundry, whether it’s OpenAI’s GPT-5.5, Anthropic’s Claude, or an open-source model. Microsoft controls the hosting, the security layer, and the data handling. The model is the intelligence engine, not the data keeper.
For most businesses, that reassurance is enough. But it isn’t the full story for everyone.
What About Regulated and Defence-Adjacent Businesses?
If your business operates under a compliance framework like DISP (Defence Industry Security Program) or requires an IRAP (Information Security Registered Assessors Program) assessment, the conversation goes further than where your data is hosted.
These frameworks are designed for businesses that handle sensitive government, defence, or critical infrastructure information.
Under DISP, businesses engaging in defence tenders are assessed across personnel, physical, cyber, and supply chain security domains, and for protected, secret, or higher-classified workloads, public cloud AI tools generally aren’t on the table at all.
IRAP assessments evaluate cloud services against the Australian Government Information Security Manual (ISM), and Microsoft completed updated IRAP assessments for Azure, Dynamics 365 and Microsoft 365 at the “Protected” level in early 2026, supporting Australian government workloads. That’s a meaningful credential for those environments.
But here’s what matters: IRAP certification covers the Microsoft platform, not every model that runs on it. A Chinese-developed model running inside an IRAP-assessed Azure tenancy doesn’t automatically inherit that certification for your workload.
The provenance of the model weights, the training data used, and the potential for embedded vulnerabilities are separate questions that your risk and compliance team would need to work through.
Independent security evaluations add weight to those concerns. In September 2025, the Center for AI Standards and Innovation at NIST evaluated several DeepSeek models and found significant shortcomings.
Agents built on DeepSeek’s R1 model were, on average, 12 times more likely than evaluated US frontier models to follow malicious instructions in agent hijacking scenarios. The same evaluation found that DeepSeek’s most tested model responded to 94% of overtly malicious requests when a common jailbreaking technique was used, compared with 8% for US reference models.
That’s a meaningful gap for any environment where agents operate with real permissions across business systems.
What this means in practice
None of this means DeepSeek-on-Azure is automatically unsuitable for regulated businesses. It means the decision requires more than a hosting check. A few questions worth asking before any model option goes near sensitive workloads:
The good news is that Microsoft’s multi-model platform is designed to give businesses that choice.
You don’t have to use DeepSeek if it doesn’t suit your risk profile. GPT-5.5, Anthropic’s Claude, and Microsoft’s own MAI models are all available within the same Copilot environment, and Microsoft’s admin tooling lets you control which models are accessible to your users.
This Is Part of a Bigger Shift
The DeepSeek story is really just one thread in a larger move Microsoft has been making through 2026. As I wrote when covering Microsoft Build in June, the company is explicitly building a multi-model future where different models serve different jobs.
Microsoft’s official position, published on 16 June, is that both Microsoft 365 Copilot and GitHub Copilot are “model-diverse by design”, meaning the platform won’t lock you into a single provider. GPT-5.5 Instant, Claude Opus 4.8 and other models already run inside Copilot Chat and Copilot Studio today.
The addition of a lower-cost open-source option for agent tasks follows that same logic: match the model to the job and the cost to the value.
This is actually how mature technology platforms tend to work. Your business probably already uses different cloud services for different workloads.
You don’t host everything on the most expensive option when a cheaper one does the job just as well. AI is heading in the same direction, though with more compliance dimensions to manage.
What It Means for Your AI Costs
If you’re running Microsoft 365 Copilot now, or planning to adopt Copilot Cowork as it scales, here’s what this shift means in practical terms.
Routine tasks will likely get cheaper to run
The tasks that Copilot Cowork handles most often, like summarising documents, drafting standard communications, and pulling data from Microsoft 365, don’t always need the most powerful model available.
Routing those to a lower-cost model while reserving frontier models for complex reasoning, legal review, or strategic analysis is a sensible design. Microsoft is building that routing logic into the platform.
Your data protection doesn’t change
When Microsoft runs any model on Azure, your business data stays inside the same Microsoft 365 security perimeter it’s always been in. Permissions, sensitivity labels, and compliance policies still apply.
The model processes the request; it doesn’t store your data, and it doesn’t share it externally. I covered the governance layer in detail when we looked at securing AI agents with Entra Agent ID and Purview, and that framework applies here too.
Cost visibility is becoming non-negotiable
One thing the DeepSeek-in-Copilot story makes clear is that AI costs are no longer just a subscription line item. As agents run more tasks autonomously, usage-based charges will show up more frequently.
Microsoft introduced a Cost Management Dashboard in the Microsoft 365 Admin Centre as part of the Work IQ API general availability, and it’s worth getting familiar with that tooling before your agent usage scales.
How to Think About Open-Source Models in Your AI Strategy
The conversation around open-source AI models is often framed as “open source versus Microsoft”, but that’s not really how it plays out in practice for most businesses. At CG TECH, I talk about this a lot with clients who are designing an AI operating model that works across multiple tools and providers.
Open-source models like DeepSeek aren’t an alternative to Microsoft. They’re one option within a broader portfolio that Microsoft itself is curating.
The more useful question isn’t “should we use open-source models?” It’s “which tasks are worth paying frontier model prices for, and which ones aren’t?”
Some good places to start asking that question:
The businesses I see getting the most from AI right now aren’t the ones using the most expensive tools. They’re the ones being deliberate about matching the right model to the right task, and knowing which tasks require a higher bar.
Where Things Are Heading
Microsoft hasn’t confirmed a release date for the DeepSeek V4 option in Copilot Cowork; the June reporting describes it as being tested, with availability expected in the “coming weeks.” It’s worth watching, because the pricing implications for businesses running agents at scale could be meaningful.
What’s already confirmed is the direction of travel: Microsoft 365 Copilot is a multi-model platform, model costs vary significantly across providers, and Microsoft is building the tooling to help businesses govern and manage that spend. The DeepSeek integration, if it lands, will be the most visible example yet of how far that strategy has come, and it’ll be optional, which matters.
For businesses operating in regulated industries, the message isn’t “avoid DeepSeek.” It’s “make the decision deliberately, with your compliance obligations in front of you.” That’s a different conversation to the one most AI vendors want to have.
If you’d like to talk through how your Microsoft 365 environment is set up for the shift to usage-based agent pricing, or what a practical multi-model strategy looks like for your business, reach out to the CG TECH team. It’s a conversation worth having before the bill, or the audit, surprises you.
About the Author
Carlos Garcia is the Founder and Managing Director of CG TECH, where he leads enterprise digital transformation projects across Australia.
With deep experience in business process automation, Microsoft 365, and AI-powered workplace solutions, Carlos has helped businesses in government, healthcare, and enterprise sectors streamline workflows and improve efficiency.
He holds Microsoft certifications in Power Platform and Azure and regularly shares practichttps://www.linkedin.com/in/carlosgarciajr/al guidance on Copilot readiness, data strategy, and AI adoption.
Sources
Recent Posts
Popular Categories
Archives