Salesforce’s AI buildout is now showing up in margin guidance. According to The Register, Mike Spencer, Salesforce’s deputy chief financial officer and head of finance, told Deutsche Bank’s Technology Conference last week that the company’s spending on Anthropic’s Claude tokens was one reason it did not raise its full-year margin outlook. The Register reports that Salesforce’s recent results showed a 20.5% operating margin under accounting rules for its fiscal second quarter, which ended July 31. Its full-year guidance, however, stood at 20.1%. Spencer attributed the gap in part to the cost of running Claude across Salesforce’s research-and-development work. The mechanism is straightforward: Salesforce has been using Claude inside its own product-development cycle, and model usage is metered through tokens. Spencer said Salesforce had “unleashed” Claude in R&D roughly six months earlier to see how much faster teams could advance the product roadmap, while keeping the option to reduce usage if costs ran too high, The Register reported. The company’s posture is now shifting from broad deployment to optimization. Spencer said Salesforce is moving into what he described as a more refined mode of model choice, where the company matches the model to the task instead of defaulting to the newest and most capable system for everything. That distinction matters for software teams. According to The Register’s account of Spencer’s remarks, Salesforce still sees cases in software development and broader technical work where newer models may be necessary. But Spencer said many tasks can be handled by older or less expensive models, including second- or third-generation systems. Salesforce is also comparing vendors internally. Spencer said the company uses OpenAI, Cursor and Claude, and is beginning to experiment with X’s Grok, The Register reported. His point was not only that the tools differ in capability, but that they carry different cost structures — a practical issue for any company putting generative AI into production workflows. The Anthropic relationship is not small. The Register also reported that Salesforce CEO Marc Benioff said in May, on the All-In podcast, that Salesforce had expected to spend $300 million with Anthropic in 2026. Salesforce and Anthropic have also launched a new partnership, according to the report. The broader signal is that AI adoption is moving from experimentation to cost control. The Register notes that Cockroach Labs CEO Spencer Kimball told the outlet in June that his company was using more open-source AI models internally as it worked to optimize AI spending. The report also cites Gartner’s finding that, in some parts of the world, spending on AI coding agents could soon exceed the salary cost of a developer. Who benefits: Anthropic benefits from Salesforce’s heavy Claude usage and the reported expectation of substantial 2026 spend. Vendors with lower-cost or task-specialized models may also benefit if large customers keep moving toward model selection by workload. Who's exposed: Salesforce is exposed to the cost of broad internal AI use if productivity gains do not offset token spending. Companies using coding agents heavily face the same pressure, especially where AI tool costs approach or exceed labor costs for some developers, as Gartner’s finding cited by The Register suggests.