Salesforce CEO Marc Benioff said on the All-In podcast that his company expects to spend roughly $300 million on Anthropic API tokens in 2026, with nearly all of that consumption directed at coding workloads. The figure, disclosed in a recorded conversation that surfaced this weekend, is one of the largest single-customer token commitments any enterprise software vendor has publicly attached to a frontier-model lab.
A token bill the size of a mid-cap SaaS budget
Benioff framed the spend as operational rather than experimental. "AI coding agents [are] awesome," he told the hosts, citing "unprecedented efficiency gains at Salesforce across service, support, distribution, and marketing." The $300 million figure is internal Salesforce consumption — Claude calls made by Salesforce's own engineers and internal agentic workflows — and is distinct from Claude-powered features that Salesforce resells to customers.
The scale matters because it reframes coding assistants from per-seat tooling into a metered cloud cost. At roughly $25 million a month of inference spend, Salesforce is now running a token-budget P&L of the kind previously associated with hyperscalers, not application vendors.
The equity side of the relationship
Salesforce Ventures has put more than $300 million into Anthropic since participating in the Series C in early 2023, and Benioff pegged the resulting stake at about 1% of a company he said is valued at roughly $380 billion. That puts the equity carry alongside the operating spend: Salesforce is both Anthropic's customer and a minority owner, with paper gains on the investment offsetting some of the API expense.
The coding focus tracks Anthropic's revenue mix. Claude has dominated agentic-coding deployments since Opus 4.x shipped, and large enterprise consumers like Salesforce are now the load-bearing customers behind Anthropic's reported $30 billion annualized run rate.
Routing layer, Slack, and the Agentforce numbers
Benioff also signaled a builder concern that will resonate with anyone running a production LLM stack: not every request needs a frontier model. He argued for an "intermediary layer" that routes complex reasoning to Claude while sending simpler calls to smaller, cheaper models — effectively endorsing model-gateway and router products at the Salesforce-scale tier.
He teased coding inside Slack — "You're going to see some cool stuff with Slack and code I'm not ready to talk about yet" — and confirmed that starting this summer, all new Salesforce customers will have Slack auto-provisioned with AI enabled. Slack revenue is expected to clear $3 billion in 2026.
Agentforce, Salesforce's agent platform, now sits at $800 million ARR, up 169% year-over-year across roughly 29,000 closed deals. Internally, agent-driven productivity has shrunk Salesforce's support headcount from 9,000 to 5,000 since the August disclosure.
What changes for builders
For practitioners, the takeaway is the cost signal. A $300M token bill from a single application vendor — almost entirely on coding — confirms that frontier-model inference is now a line item large enough to justify dedicated routing infrastructure, caching, and small-model fallbacks. Benioff's public ask for a routing layer is, in effect, a procurement spec.
