Microsoft plans to replace Copilot Cowork with DeepSeek V4

Microsoft may be planning to perform a “heart transplant” on its flagship AI agent product using DeepSeek.

According to a June 17 report from Axios, Charles Lamanna, Executive Vice President of Copilot, AI Agents, and Platforms at Microsoft, revealed that the company is switching Copilot Cowork’s billing model to “compute-based charging” and is also evaluating replacing the current backend Anthropic and OpenAI models with a fine-tuned version of DeepSeek V4 (or other open-source models). Microsoft’s stated timeline is—an official announcement of the final choice within a few weeks.

The most interesting part is not “Microsoft embracing open source again,” but the fact that even a giant that sells cloud services, holds 49% of OpenAI's revenue rights, and has just renewed the Claude on Azure agreement with Anthropic is beginning to find the tokens of closed-source models too expensive.

The Agent Economy Is Being Undermined by Its Own Bills

To understand why Microsoft is considering this move, you first need to know what Copilot Cowork is.

It’s not last year’s lightweight Copilot that “helps polish text in Word,” but rather an autonomous task-executing AI agent like Claude Code or OpenAI Codex. You give it a goal, and it decomposes the tasks, writes code, runs tools, fixes bugs, and loops through multiple rounds of execution.

The problem lies in this “runs itself” feature.

In traditional Chat mode, one conversation typically consumes at most thousands to tens of thousands of tokens. But agents are a different beast—an even moderately complex coding task can run millions of tokens worth of context in the background, making dozens or even hundreds of model calls. Every step of reasoning, every tool call with backfill, every round of self-reflection—they all burn money.

Lamanna used a not-so-diplomatic phrase: agents bring productivity, but also “crazy AI bills.”

This is as close to a complaint as it gets. For example, Anthropic’s latest flagship Fable 5 costs $50 per million tokens—and that’s enterprise pricing; meanwhile, DeepSeek V4 Pro with a permanent 75% discount costs $0.87 per million tokens. That’s about a 57x difference.

For an agent that needs tens of millions of tokens to complete a single enterprise process, this isn’t about saving some money—it’s about whether commercialization is even possible.

Microsoft’s Calculation: Switching from Subscription to Usage-Based

Microsoft’s move actually combines two changes, though the public tends to only notice the model replacement.

Step one is billing model modification. Copilot Cowork previously used a subscription model, charging per seat. The new model is usage-based, charging based on the actual compute burned by the customer. This change indicates that the old per-user packaging no longer covers the costs for heavy users—especially development teams and enterprise automation scenarios that truly put agents to work.

Step two is replacing the underlying model. Once billing becomes transparent, model costs will directly pass to customers. If Claude or GPT series continue to be used, customers will clearly see bills skyrocketing and will resist usage. Microsoft must give enterprises a “cheap but capable” default option.

And DeepSeek V4 is now on the shortlist.

Where DeepSeek V4 Excels

It’s worth explaining the V4 generation clearly, because it’s very different from the “V3 era DeepSeek” many people remember.

On April 24 this year, DeepSeek released the V4 preview and simultaneously open-sourced it, with two main versions:

• DeepSeek-V4-Pro: Flagship version, benchmarked against top closed-source models like Opus 4.6
• DeepSeek-V4-Flash: Lightweight version, aimed at large-scale, accessible scenarios

Architecturally, V4 didn’t simply pile on parameters—it introduced a triple combo:

• Hybrid Attention Architecture (CSA + HCA)
• Manifold-Constrained Hyper-Connections (mHC)
• Muon Optimizer

The performance is fierce. In 1 million-token context scenarios, V4-Pro’s per-token inference compute is only 27% of V3.2, KV cache memory usage drops to 10%. This means the same GPU can handle several times the concurrent inference tasks.

More critically, there’s Agentic ability. In Agentic Coding benchmarks, V4-Pro has reached the best level among open-source models, matching Opus 4.6 on some metrics. In other words, for Copilot Cowork’s coding and enterprise process automation use cases, V4 isn’t just “good enough”—it’s genuinely capable.

Add another layer: Fully open-sourced under the MIT license, supporting local deployment and secondary development.

These two features combined make it almost tailor-made for Microsoft—they can take the model weights, fine-tune them on Azure specifically for Copilot Cowork workflows, and run the whole thing entirely on their own cloud, avoiding any inference fees to third parties.

Microsoft Makes Data Security Promises in Advance

Aware that “using a Chinese open-source model” might raise political and compliance concerns among enterprise clients, Microsoft is careful with its wording.

Lamanna made clear promises:

• Model fully hosted on Azure
• Customer data retained in Microsoft’s cloud
• Subject to Azure’s enterprise security, compliance, and data residency controls

In plain terms: We only take the weights and won’t let any tokens leave Azure’s boundaries.

This language mirrors what Azure used when integrating Llama and Mistral. Essentially, they treat “open-source models” as assets that can be locally deployed, not as external services. For clients in regulated industries (finance, healthcare, government), this distinction matters greatly.

What This Means for OpenAI and Anthropic

Not great.

Microsoft is OpenAI’s largest backer and biggest distribution channel, with Copilot products being the GPT series’ largest enterprise traffic source. If Copilot Cowork swaps out the underlying models, OpenAI will be sidelined in one of the most lucrative enterprise AI agent tracks by its foremost partner.

Anthropic’s position is more awkward. The Claude on Azure agreement was critical for its enterprise market expansion, but now Microsoft is publicly saying: Claude is too expensive, we’re looking for a replacement.

Behind this lies a bigger trend—the capability gap between models is narrowing, while the price gap is widening.

Since 2026, major closed-source providers have generally raised token pricing (Fable 5, GPT-5.5 have increased rates), with the logic being “I’m ahead in capability, so I’m expensive.” But when open-source models (DeepSeek, Qwen, GLM, etc.) narrow the Agentic ability gap to within 10–20%, for most enterprise scenarios, saving 50x in costs is far more attractive than gaining 10% capability.

That’s the calculation Lamanna and team are making.

The Fine-Tuned Version Is the Key Variable

One detail overlooked by many: Microsoft isn’t directly using DeepSeek V4, but “a fine-tuned version of DeepSeek V4.”

This is more significant than it sounds.

General V4-Pro already excels at Agentic Coding, but Copilot Cowork’s scenarios involve not just coding, but also:

• Microsoft Graph API calls
• Office document operations
• Teams workflows
• Azure DevOps integration
• Enterprise identity and permission chains

These tool calls and long-chain reasoning require specific instruction fine-tuning and tool-use training. Microsoft can take V4’s open-source weights, feed in large amounts of their tool usage-path data, and produce a “uses Microsoft’s ecosystem only, but does it smoothly” specialized version.

The benefits: specialized capabilities, controllable inference costs, and keeping ecosystem moats inside Microsoft. The downside: diverging from DeepSeek’s official version means redoing fine-tuning for future V4.5 or V5 releases.

But for a company of Microsoft’s scale, this MLOps cost is negligible.

Domestic Developers’ Perspectives

If you use Copilot Cowork or similar agents, several points from this change are worth noting:

1. Prices will drop significantly. Usage-based pricing combined with cheaper models means heavy users’ marginal costs could fall by an order of magnitude. Scenarios previously avoided due to billing concerns can be re-evaluated.

2. Capabilities may fluctuate temporarily. Switching from Claude/GPT to a fine-tuned DeepSeek may cause regressions in certain extreme long-chain reasoning or rare language contexts. Key workflows should undergo A/B testing.

3. Multi-model choice will become the norm. Microsoft’s statements suggest Copilot products will likely evolve into architectures routing tasks to different models—cheap tasks through DeepSeek, high-difficulty tasks through Claude/GPT. This aligns with many AI gateway product strategies.

By the way, OpenAI Hub (openai-hub.com) has already integrated the full DeepSeek V4 lineup (Pro and Flash), and also supports GPT, Claude, Gemini, and other mainstream models. One key allows switching calls. This means small teams can replicate Microsoft’s “multi-model cost-routing” approach without applying to each provider individually or handling domestic access issues—convenient for agent cost optimization.

A Judgment

The true signal from Microsoft’s move isn’t some nationalist narrative about “Chinese models winning,” but that agent commercialization is being throttled by cost structures.

Agents consume tokens on a scale unlike traditional chat. When a product relies on repeated model runs to complete tasks, the marginal cost of the underlying model directly determines whether it can scale. Under these constraints, “good enough and cheap” open-source models will increasingly be pushed to the core of B2B products, while closed-source flagships will be relegated to “high-difficulty task fallback.”

Claude Code and Codex won’t follow suit in the short term—their brand positioning precludes a turnaround. But Copilot Cowork is Microsoft’s own product line, with much greater flexibility.

The moment Microsoft makes its official announcement in a few weeks could be one of the most important turning points for the enterprise AI market in 2026.

---

References

• Anthropic and OpenAI models are too expensive — Microsoft’s AI agent may switch to a fine-tuned DeepSeek V4 - IT Home — Axios report summary on Microsoft Copilot Cowork adopting DeepSeek V4