Here‘s the headline number: Anthropic’s Fable 5 charges $50 per million output tokens. DeepSeek V4 Pro charges $0.87 per million output tokens. That‘s a 57x difference .
For most routine enterprise tasks — drafting emails, managing inboxes, summarizing documents — the cheaper model does the job just as well. Microsoft’s own testing suggests open-source models can accomplish the same tasks with over 90 percent cost reduction .
Microsoft is considering a switch. Not because OpenAI or Anthropic are bad — because they‘re expensive. And when you’re running a product like Copilot Cowork, where some users execute hundreds of tasks per week, the cost differential is existential.
“The cost can get very high,” Microsoft executive Charles Lamanna said . The company is moving Copilot Cowork to token-based pricing and actively evaluating DeepSeek V4 as a lower-cost alternative .
The Three Numbers That Tell the Real Story
130.7 billion. OpenAI‘s 2025 revenue . That’s up from $3.7 billion in 2024 — a 253% increase. Revenue is growing fast. But cost is growing faster.
340 billion. OpenAI‘s 2025 total costs and expenses . That’s $34 billion in spending on $13 billion in revenue. For every dollar OpenAI earned in 2025, it spent $2.60 .
209.2 billion. OpenAI‘s operating loss in 2025 . Even after stripping out non-cash accounting adjustments — which made the reported $38.5 billion net loss look even worse — the cash operating loss was roughly $8 billion. That’s still eight billion dollars of cash burned in a single year.
The numbers point to an uncomfortable question: if even your closest partner starts flinching at the cost, how sustainable is the model?
Why Microsoft’s Move Matters More Than the Headline
Microsoft isn‘t some random enterprise customer. It’s OpenAI‘s largest investor. It’s OpenAI‘s exclusive cloud provider. In 2025 alone, OpenAI paid Microsoft $17.2 billion for cloud services and model training .
When your biggest partner — the one that literally hosts your infrastructure — starts shopping around for cheaper alternatives, that’s not a negotiation tactic. That‘s a signal.
The shift to usage-based pricing for Copilot Cowork is itself telling. The product was previously offered as an unlimited flat-fee service. Microsoft discovered that power users were consuming resources at rates that made the economics impossible. The company’s response: charge for what you use, and consider switching to models that cost 50x less per token .
The Market Is Already Moving
Enterprise customers are voting with their wallets, and the direction is clear. Vercel‘s data shows DeepSeek’s share of US enterprise AI usage jumped from 1 percent in April to 17 percent in May .
AI agent startup Lindy told The Wall Street Journal that switching from Anthropic to DeepSeek saved “millions of dollars.” The company moved 100 percent of its user-facing traffic to DeepSeek and found performance on core tasks like email management matched or exceeded Anthropic‘s Sonnet model .
OpenRouter, the largest AI API aggregator, reports that a combination of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro beats GPT-5.5 and Opus 4.8 on benchmarks at about half the cost — closing the gap to Fable 5 to under 1 percent .
This is not an anomaly. It’s a structural shift. If cheaper models can deliver 99 percent of the performance at 2 percent of the cost, the pricing power of premium AI models is eroding faster than anyone expected.
The Question Worth Asking
Microsoft‘s shift to usage-based pricing and its evaluation of DeepSeek is the immediate story. But the deeper question is about what happens to the trillion-dollar AI infrastructure build-out if enterprise customers stop paying premium prices for premium models.
Microsoft’s move is the first major sign that cost pressure is real enough to overcome even the closest partnership. And when your biggest customer starts looking at the math, the math tends to win.
P.S. The difference between DeepSeek and Fable 5 on simple enterprise tasks isn‘t quality. It’s price. Microsoft didn‘t discover that OpenAI was bad — it discovered that good enough is a lot cheaper than best. And when the gap is 57x, “good enough” starts looking very attractive.
