For Anthropic, one piece of news brings both good and bad tidings: two heavyweight customers, Microsoft (NASDAQ: MSFT) and Meta, are working to reduce the costs generated by employee use of Claude.

The negative side is that Microsoft's internal annualized spending on Claude has fallen by more than 33% from a level that peaked above $1 billion earlier this year.

The positive side is that Microsoft's own customers are increasing their spending on Claude. Microsoft continues to boost payments to Anthropic to cover customer usage generated by Claude-powered features inside Microsoft's Copilot AI tools, and that revenue roughly offsets the reduction in Microsoft's internal procurement costs.

Now Microsoft is signaling that it wants to help customers replicate what it has achieved internally 鈥?replacing Claude inside Copilot with cheaper models without lowering quality. That way, Microsoft can keep customer costs lower than if those customers used Claude directly on their own.

Microsoft charges Copilot users a base fee of $30 per seat per month; advanced features such as AI coding assistants and intelligent agents are billed separately, with the amount depending on how much AI model consumption a customer incurs through the software. In theory, if Microsoft can complete the same tasks with lower-cost AI, those add-on fees would fall.

Microsoft Executive Vice President Jacob Andreou said: "Copilot has in fact already incorporated a wide variety of models. Our core responsibility is to control quality, ensure the final output, and also manage efficiency for customers, striving to deliver good results at the right cost."

Background review

Over the past year or so, Microsoft has relied on Anthropic technology as well as other AI models to support the GitHub Copilot coding tool and some complex functions in Office 365 Copilot, such as generating Excel formulas and producing PowerPoint presentations. But more recently Microsoft has begun bringing in OpenAI models such as GPT-5.6 Sol to handle advanced tasks. Under the commercial agreement between the two sides, Microsoft does not pay extra to use these OpenAI models. Microsoft is also using its in-house MAI models to replace Claude in some coding scenarios in order to cut costs.

Microsoft is further optimizing how Copilot executes multi-step tasks. For example, when AI generates a PowerPoint presentation, it originally needed to write roughly 100,000 characters of general underlying code to build the slides. Microsoft has made programming optimizations to Copilot so that this script can be reused directly each time a presentation is generated, no longer incurring additional token charges. Andreou said: "We do not need to generate the script from scratch, and we do not need to consume large amounts of frontier large-model tokens to write this code. We can both improve output quality and significantly reduce costs."

Whether these Microsoft moves can significantly lower customers' Claude bills is still uncertain; this does not mean Anthropic's models will immediately disappear from Copilot. Andreou noted that Anthropic models are still the best fit for some specific tasks, and Microsoft's top priority has always been the output quality of Copilot. "Our core goal is to provide an extremely high-quality experience."

What comes after the SaaS apocalypse?

Earlier, the market worried that AI would completely replace core business record systems, but this extreme "SaaSpocalypse" panic has cooled. Yet traditional enterprise software like Workday (NASDAQ: WDAY) is facing a subtle but existential threat: being reduced to "unintelligent underlying infrastructure." Workday's core human resources and finance databases remain indispensable for auditable payroll and employee record storage. But customers are increasingly bypassing Workday's native interface. Companies like NinjaOne and Druva are no longer buying Workday's built-in proprietary AI tools; instead they deploy external AI agents from Anthropic, Microsoft and others, pulling Workday data directly through API interfaces. This shift strips away Workday's high-value data analysis layer and degrades its software into a back-end basic service that is merely called upon, with users no longer actively logging into its system.

The report noted that despite heavy promotion, commercialization of Workday's self-developed AI features got off to a poor start. In other words, traditional enterprise software vendors are likely next to increase discounts on their self-developed AI features, allow customers to sign shorter-term contracts, and even commit in writing that AI features will meet specified performance standards. Many vendors have already begun doing so.