Salesforce Anthropic token spend set for $300M in 2026, mostly coding

Salesforce Anthropic token spend is moving from a headline number to something more revealing: a sign that AI is becoming a core operating cost inside major software companies. Marc 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.

That detail matters because it shifts the story away from flashy demos and toward day-to-day economics. Benioff is not talking about a small pilot. Instead, he is describing a level of AI usage that suggests coding agents are being treated as infrastructure.

He also tied that spending to a broader product push. Salesforce is working on technology to make coding easier inside Slack, extending the company’s bet that workplace collaboration and AI agents will increasingly blur together.

Salesforce Anthropic token spend points to a bigger AI cost shift

Benioff’s projection, shared on the podcast, puts a price tag on how deeply Salesforce appears willing to embed Claude into its internal workflows. He said most of the expected $300 million token spend will go toward coding, a striking signal for an enterprise software company whose own products are built by large engineering teams.

The implication is simple: Salesforce believes AI coding agents can lower the cost of building software enough to justify a massive usage bill. Benioff framed those tools as a way to make everything at Salesforce cheaper to build.

This is also a useful marker for the broader enterprise AI market. Token-based pricing has often sounded abstract outside technical circles, but a projected spend of this size makes the model easy to understand. For large customers, AI is no longer just a software subscription. It is becoming a metered utility.

Why Slack is central to Marc Benioff’s coding plan

Slack sits near the center of that strategy. Benioff said Salesforce is building technology to make coding easier inside the messaging platform, hinting at tighter connections between developer work and the place where teams already communicate.

That builds on changes Salesforce already made in March, when Slack was overhauled with more than 30 new AI capabilities for Slackbot. Those additions pushed Slackbot beyond a chat assistant and into a more active role inside workplace workflows.

All of those new Slack capabilities run on Anthropic’s Claude. From this summer, every new Salesforce customer will also have Slack automatically provisioned and AI-enabled from day one. That is a significant distribution move. Rather than asking customers to adopt Slack AI as an add-on, Salesforce is making it part of the default setup for new accounts.

In practice, that means AI features arrive pre-enabled inside a collaboration product. As a result, usage tends to move closer to routine behavior. That can increase token consumption quickly, especially if coding, task execution, and internal support all start flowing through the same interface.

The cost case behind the rollout

Benioff said AI agents have already delivered “unprecedented” efficiency gains at Salesforce across service, support, distribution, and marketing. His latest comments suggest the company now sees engineering as the next major area where those gains can compound.

That helps explain the Salesforce Anthropic token spend. 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.

Benioff also pointed to a second layer of strategy: cost control. He argued that not every task should be sent to a frontier model. Instead, he described an “intermediary layer” that could route simpler work to cheaper models and reserve Claude for harder reasoning jobs.

That idea is more important than it may sound. At a projected annual spend of $300 million, even modest routing improvements could have a meaningful financial effect. It also reflects a maturing enterprise AI approach: not just buying model access, but managing model usage the way companies manage cloud infrastructure.

A new enterprise AI spending model

There is a broader signal here for investors, software buyers, and rivals. Enterprise AI costs are shifting from experimentation budgets into core operating expenses. Once companies begin to rely on AI for coding and daily workflows, token usage starts to resemble a recurring utility bill rather than a temporary innovation project.

Salesforce’s push also suggests the next competitive edge may not come only from picking the best model. It may come from deciding when to use the best model, when to use a cheaper one, and how seamlessly that choice is hidden from employees.

Salesforce’s wider Anthropic relationship

The operational partnership sits alongside a financial one. Salesforce has invested more than $300 million in Anthropic and holds about a 1% stake, according to the information discussed around Benioff’s comments.

That means the relationship runs on two tracks at once: Salesforce is a major user of Anthropic’s technology through Claude-powered products and internal coding use cases, and it is also an investor with a direct stake in Anthropic’s upside.

This combination gives the Salesforce Anthropic token spend extra weight. The company is not simply paying a vendor. It is backing a model provider financially while building product experiences around that provider’s technology.

Benioff’s comments also add context to how Salesforce got here. He has said Microsoft blocked Salesforce from investing in OpenAI, pushing the company toward Anthropic instead. Whatever the backstory, Salesforce now appears to be making Anthropic central to both its AI product layer and its internal productivity bet.

What changes next for customers and the market

For customers, the near-term change is clear: Slack will be present earlier, with AI turned on from the start for new Salesforce accounts beginning this summer. That could make Slack less of a separate collaboration app and more of a front door for Salesforce’s broader AI experience.

For the market, the bigger takeaway is that Claude coding agents and Slack-native AI are being positioned as tools that can justify very large enterprise AI costs if they lower labor and development expenses enough. Salesforce is effectively testing whether token spend at this scale can be absorbed as a normal cost of building and running a modern software company.

If that math holds, other enterprise players may follow with their own versions of the same playbook: one collaboration layer, one routing layer, and a very large model bill underneath it.