Benioff: Salesforce to Spend $300M on Anthropic Tokens

Salesforce CEO Marc Benioff expects his company to spend $300 million on Anthropic tokens this year, almost entirely on coding. The announcement came in an episode of the All-In podcast published on Friday, marking one of the clearest signals to date that enterprise AI usage has moved from pilot programs to core operational infrastructure. Benioff told the podcast that AI coding agents are driving unprecedented efficiency and teased new Slack coding tools.

Benioff discusses AI infrastructure investment on podcast with futuristic AI brain visualization and digital workspace

AI Token Spend Becomes Core Enterprise Cost

Benioff said on the All-In podcast published Friday that Salesforce expects to spend $300 million on Anthropic tokens in 2026, with most of that bill tied to coding. The figure represents a fundamental shift in how major software companies view artificial intelligence. Rather than treating AI as an experimental feature, Salesforce has integrated Anthropic’s Claude models into daily workflows at a scale that transforms token consumption into a recurring utility expense.

Speaking on the All-In Podcast, Benioff said coding agents have unlocked capabilities that were previously out of reach, including faster software deployment, AI-driven outbound prospecting and the ability to implement and sell products at the same time. Benioff was emphatic in his praise, stating in the podcast that “these coding agents are awesome” and calling Anthropic “awesome.”

If coding becomes the biggest use case, then the company is effectively betting that more tokens can mean fewer bottlenecks, faster iteration, and lower development costs. For an enterprise software company whose products are built by thousands of engineers, the investment signals confidence that AI-assisted coding can deliver ROI sufficient to justify hundreds of millions in annual spending.

Slack Coding Integration on the Horizon

With the productivity boosts he’s seen with coding agents, Benioff revealed in the podcast that Salesforce is also working on ways to make coding easier in Slack, the productivity platform Salesforce has owned since 2021. “We’re even working on technology inside Slack to make it easier for everybody to code,” he said. “You’re going to see some cool stuff with Slack and code I’m not ready to talk about yet.”

Salesforce overhauled Slack in March, unveiling more than 30 new AI capabilities for Slackbot that transform it from a conversational assistant into an agentic system capable of transcribing meetings, monitoring desktop activity, executing tasks through third-party tools via the Model Context Protocol, and functioning as a lightweight CRM. All of the new capabilities run on Anthropic’s Claude.

Benioff has described Slack as “the interface to AI” and noted that AI companies including OpenAI and Anthropic run their own operations on the platform. Slack revenue is expected to hit $3 billion this year, and Salesforce’s Agentforce business, its dedicated AI agent product line, reached $800 million in annual recurring revenue as of the most recent earnings, up 169% year-on-year with 29,000 deals closed.

Cost Control Through Model Routing

While announcing the massive token expenditure, Benioff also outlined a strategy for controlling costs. Benioff said that while he projects a $300 million spend on Anthropic this year, he doesn’t believe every token a company employee use has to go to a frontier model like Anthropic’s Claude. The CEO called for an “intermediary layer” in the future that could determine which token inputs should go to Anthropic and which ones may be suitable for smaller models.

Claude Opus 4.7, released this month, is priced at $5 per million input tokens and $25 per million output tokens. Smaller models from Anthropic’s own Haiku line, or from open-weight competitors like Meta’s Llama and DeepSeek, cost a fraction of that. At $300 million in annual spend, even modest routing optimisation could save Salesforce tens of millions of dollars, and Benioff appears to be signalling that Salesforce will build that optimisation rather than wait for Anthropic to offer it.

The routing strategy reflects a maturing approach to enterprise AI deployment, where companies manage model usage with the same discipline applied to cloud infrastructure spending. By reserving expensive frontier models for complex reasoning tasks while directing routine queries to cheaper alternatives, Salesforce aims to maximize efficiency across its AI operations.

Dual Financial and Operational Relationship

Salesforce has also invested more than $300 million in Anthropic as a company, beginning with its Series C round in early 2023, giving it roughly a 1% stake in a company now valued at $380 billion. Benioff has said Microsoft blocked Salesforce from investing in OpenAI, which redirected the company toward Anthropic.

The dual relationship means Salesforce operates simultaneously as a major customer consuming Anthropic’s API services and as an investor with equity exposure to the company’s growth. That investment has produced a paper return of more than ten times the original outlay. Whether the $300 million in tokens produces a comparable return in engineering productivity is the question Benioff is betting on.

Salesforce has also launched Headless 360, an API-first platform with 60+ MCP tools designed to give AI agents, including Claude Code, direct access to its enterprise stack. This infrastructure enables Claude to integrate deeply into Salesforce’s development environment and customer-facing products.

Key Facts

Salesforce expects to spend $300 million on Anthropic tokens in 2026, almost entirely on coding
Slack revenue is expected to hit $3 billion this year
Salesforce’s Agentforce reached $800 million in annual recurring revenue, up 169% year-on-year with 29,000 deals closed
Salesforce holds roughly a 1% stake in Anthropic, valued at $380 billion
From this summer, every new Salesforce customer will have Slack automatically provisioned and AI-enabled from day one
Claude Opus 4.7 costs $5 per million input tokens and $25 per million output tokens