Marc Benioff said on the All-In podcast published Friday that Salesforce expects to spend roughly $300 million on Anthropic tokens in 2026 — almost entirely on AI coding agents. That is Benioff’s own projection, not an audited figure. But the signal is clear: AI model costs have moved from pilot budgets into operating expense territory.

Key takeaways:

• Salesforce projects ~$300M in Anthropic token spend in 2026, mostly on coding.
• Benioff calls for a routing “intermediary layer” to direct low-complexity queries to cheaper models.
• Support headcount already fell from 9,000 to 5,000; same cost curve is now playing out in engineering.
• Agentforce hit $800M ARR in fiscal 2026, up 169% year-on-year across 29,000 deals.

Why does coding dominate Salesforce’s token bill?

Benioff described AI coding agents as delivering “unprecedented” efficiency — faster iteration, lower build costs, higher output volume. He teased a forthcoming Slack coding integration as the next step, though he said he was “not ready to talk about” specifics.

Anthropic’s annualized revenue run rate grew from roughly $9 billion at end of 2025 to approximately $30 billion by March 2026, largely driven by enterprise coding, legal, and financial services adoption. Salesforce’s projected spend alone represents roughly 1% of that run rate — concentration that matters when negotiating enterprise model contracts.

This follows Benioff’s existing pattern on service. In August 2025, he disclosed that AI agents had reduced Salesforce’s support headcount from 9,000 to 5,000. The same cost-savings playbook is now running in engineering. Meanwhile, Agentforce hit $800M in ARR in fiscal 2026 — the token bill is funding a business line at real revenue scale, not absorbing unrecovered losses.

What does Benioff’s “routing layer” mean for operators?

The more useful insight from Benioff’s comments was not the dollar figure — it was the caveat. He said not every token an employee generates needs to reach a frontier model like Claude. He called for an “intermediary layer” to route low-complexity inputs to smaller, cheaper models and reserve frontier capacity for high-value tasks.

That is a description of mature enterprise AI cost management. Most operators today send everything to frontier models because setup is simpler and monthly bills haven’t crossed the threshold where routing pays for itself. When they do, routing becomes a meaningful lever — and Benioff is already thinking about it at $300M scale.

What should operators do now?

The near-term move is not to match Salesforce’s token bill. It is to start treating AI model spend like other infrastructure costs: budget it, forecast it, and architect for it.

Three actions for this quarter: (1) Audit token spend across all teams — most operators lack a consolidated view. (2) Identify your top use cases by token volume and ask whether each genuinely needs a frontier model, including your AI coding tools. (3) Ask AI vendors about volume discount tiers before your bill grows large enough to weaken your negotiating position.

Operator posture: ask sharper vendor questions on model pricing tiers and contract portability now. This trend mirrors the workforce restructuring pattern AdvancedAI has tracked across the industry — the cost-efficiency story and the headcount reduction story are the same story, playing out across function after function.

Watch next: whether other enterprise software CEOs make comparable token spend disclosures in earnings calls or public statements. One CEO on a podcast is an anecdote. Three or four rewrites how boards think about AI as an operating cost.

---

Frequently asked questions

What are AI tokens and why does a $300M spend matter?

Tokens are the units of text AI models process — roughly three-quarters of a word each. Providers bill based on volume consumed. Salesforce’s $300M projection signals that AI model costs have crossed from small-scale experiments into core operating budgets for large enterprises. Operators should model token costs the same way they model cloud compute or major SaaS licenses.

Does Salesforce’s coding spend mean engineering job cuts are coming?

Benioff didn’t say directly, but the service pattern is instructive: after deploying service agents, Salesforce cut support headcount from 9,000 to 5,000. He now describes coding agents producing similar productivity gains. The pace of workforce impact will depend on agent quality improvement rates and how organizations choose to redeploy — or reduce — engineering capacity.

When does model routing make sense for a smaller enterprise?

Formal routing architecture — directing queries to different models by complexity — typically justifies its overhead when monthly AI API costs become material: roughly when they exceed your largest individual SaaS subscription. Until then, the useful first step is consolidated usage monitoring so you know which teams are driving token volume and for what tasks.

---

By Advanced AI · May 17, 2026 · Briefings