|
Welcome back!
There was good news and bad news for Anthropic in our story yesterday that showed how two of its mega-customers, Microsoft and Meta, were cutting their staffs’ Claude bills.
Among the bad news was that Microsoft had reduced its spending on Claude by more than 33% from its peak of over $1 billion, on an annualized basis, earlier this year.
The good news is Microsoft’s customers are spending more on Claude themselves. Microsoft has been steadily increasing the payments it sends onto Anthropic to cover its customers’ use of Claude-powered features in Microsoft’s Copilot AI tools, roughly offsetting the reduction in internal spending.
But now, Microsoft is signaling it wants to help its customers do the same thing it achieved internally—that is, swap out Claude for cheaper models in Copilot without sacrificing quality, which could allow them to keep costs lower for customers than if customers just used Claude. Microsoft charges Copilot users a monthly $30 per-seat fee, and additional fees for advanced features like AI coding and agents, depending on the cost of AI models they consume through the software. In theory, those additional fees would come down if Microsoft can use cheaper AI to automate the same tasks.
“[Copilot] has really become this tapestry of all these different models where our job is to own quality and to own the outcome,” said Jacob Andreou, a Microsoft executive vice president who is heavily involved in developing Copilot. “And our job's also to own efficiency for you, to try to give you a great output for the right cost.”
As a refresher, Microsoft for more than a year has relied on Anthropic’s technology to power complex tasks in its GitHub Copilot coding tool, as well as certain features in its Office 365 Copilot features, such as creating formulas in Excel or generating PowerPoint presentations. But recently the company has been swapping in OpenAI models like GPT-5.6 Sol to handle advanced tasks. Microsoft can use those models without paying OpenAI, thanks to the companies’ commercial agreement. It’s also been swapping in its homegrown MAI models to replace Claude for some coding tasks at a lower cost.
Microsoft has also been optimizing how Copilot carries out certain steps in those multi-step tasks. For instance, every time AI creates a PowerPoint presentation, it would typically write about 100,000 characters of generic code needed to create the slide deck. But Microsoft programmed Copilot to paste in that script every time it creates a PowerPoint deck so generating the code doesn’t contribute to customers’ token bill.
“We don't need to create that script from scratch and spend a whole bunch of frontier tokens to create that script,” Andreou said. “We can actually improve the quality while also reducing the cost a ton.”
Whether those efforts by Microsoft will make a meaningful dent in customers’ Claude bills is unclear, and it doesn’t mean Anthropic’s models are vanishing from Copilot any time soon; Andreou said Anthropic’s models are still the best-suited to certain tasks and that Microsoft’s first priority is the quality of Copilot’s work.
“The main thing we're focused on is just delivering a super high quality experience,” he said.
What Comes After the SaaSpocalypse
While earlier fears of a total “SaaSpocalypse” replacing core systems of record have faded, traditional enterprise applications like Workday face a subtle but existential threat: becoming “dumb infrastructure.”
Workday’s critical HR and finance databases remain indispensable for running audited payroll and housing employee records, but customers are increasingly bypassing its user interface, my colleague Laura Bratton reported today. Instead of buying Workday’s proprietary in-app AI tools, companies such as NinjaOne and Druva are deploying external AI agents from providers like Anthropic and Microsoft to pull data from Workday via application programming interfaces.
This shift threatens to strip Workday of its high-value analytics layer, turning its software into a back-end utility that users call rather than visit. Meanwhile, Workday’s efforts to monetize its own AI features have stumbled out of the gate despite aggressive promotion, the report showed. In other words, don’t be surprised if traditional enterprise app firms boost the discounting of their own AI features, allow customers to sign shorter contracts, or give customers written guarantees that AI features in those apps will perform at a certain level. Many have already started to do those things, as Laura previously reported.
|
