A Record-Scale Commitment to Anthropic Tokens
Salesforce CEO Marc Benioff has revealed plans to spend USD 300 million (approx. RM1.38 billion) on Anthropic tokens in a single year, a scale of enterprise AI spending that stands out even in a rapidly heating market. Tokens are the basic billing unit for Anthropic’s Claude models, representing the text processed during coding and other tasks. Benioff says this spend will be “almost entirely for coding,” effectively turning Claude into a core development engine for Salesforce’s software stack. The company is already one of Anthropic’s larger commercial customers and previously invested over USD 300 million (approx. RM1.38 billion) for about a 1% equity stake. This move underscores a strategic shift from AI as a peripheral feature to AI as a primary production resource, with Salesforce treating model access and AI token costs as a major line item in its engineering budget.
From Coders to AI Supervisors: The New Engineering Model
Alongside its unprecedented AI token purchases, Salesforce has frozen software engineering hiring, redirecting how its roughly 15,000 engineers work. Benioff says AI tools, including Anthropic’s models, OpenAI Codex, Cursor, and Salesforce’s own Agentforce, have driven more than 30% productivity gains and now handle 30% to 50% of the company’s workload. Instead of writing every line of code, engineers increasingly supervise AI-generated output, review quality, and design workflows for coding agents. Salesforce is “not adding any more software engineers” in the near term, even as it plans to hire 1,000 to 2,000 salespeople to sell AI products. The company has already used AI agents to cut its support staff from 9,000 to 5,000, demonstrating tangible ROI. Engineering, in this model, becomes less about manual coding and more about orchestration, governance, and system-level design around AI-native pipelines.
Slack as a Frontline for AI Coding Tools
Salesforce is not just consuming Anthropic’s capabilities behind the scenes; it is weaving them directly into worker-facing tools. Benioff confirms that Salesforce is building Slack AI coding tools, bringing Claude-powered agents into the collaboration platform it acquired for USD 27.7 billion (approx. RM127.3 billion). While details remain under wraps, he promises “cool stuff with Slack and code,” signaling that developers may soon write, review, and deploy code within their everyday messaging environment. This integration complements Agentforce and Headless 360, Salesforce’s API-first platform that exposes more than 60 MCP tools and grants coding agents direct access to enterprise systems. By situating AI coding workflows inside Slack, Salesforce aims to turn its communication hub into a development cockpit, tightening its ecosystem and differentiating Slack from rival workplace tools with deeply embedded, enterprise-grade AI coding experiences.
AI Token Costs and the Rise of Multi-Model Orchestration
Spending USD 300 million (approx. RM1.38 billion) annually on AI tokens forces Salesforce to treat AI consumption like a critical utility. Benioff stresses that “not every token should go to a frontier model,” calling for an “intermediary layer” that routes requests intelligently. Simpler tasks can be handled by smaller, cheaper models, while Anthropic’s Claude is reserved for complex reasoning and advanced coding. This approach reflects a broader enterprise trend: multi-model orchestration and cost-aware routing as core infrastructure. Salesforce’s Headless 360 and its suite of MCP tools exemplify how enterprises are building AI-native platforms that balance capability, latency, and cost. Rather than relying on a single foundation model, Salesforce is constructing a tiered architecture where AI token costs are continuously optimized, akin to traffic engineering in cloud networks.
Outsourcing Intelligence: A New Enterprise AI Strategy
Salesforce’s Anthropic investment highlights a strategic pivot sweeping the software industry: outsourcing core AI capabilities to specialized providers instead of building everything in-house. Benioff calls Anthropic “a rocket ship that will not stop,” praising its focus on coding agents while others chase diverse consumer applications. For Salesforce, the competitive edge increasingly comes from how it embeds Claude and other models into products like Slack and Agentforce, not from owning the underlying model weights. This mirrors cloud’s evolution, where companies rent compute rather than run their own data centers. Now, they are effectively renting intelligence. As enterprise AI spending surges, vendors that can orchestrate external models, manage AI token costs, and redesign their workforce around AI-generated code oversight will gain a structural advantage over rivals still treating AI as a bolt-on feature.