Salesforce to spend $300 million on Anthropic tokens for AI coding
Salesforce will spend $300 million on Anthropic tokens in 2026, almost entirely on AI coding agents. CEO Marc Benioff believes this is cheaper than hiring engin

Salesforce will spend $300 million on Anthropic tokens in 2026, almost entirely on AI development agents. The figure was mentioned by CEO Mark Benioff on the All-In podcast.
Spending on AI as a regular budget line
$300 million a year on tokens is an investment to replace the most expensive part of development. Benioff is honest: this is not an experiment, but a business decision. Cloud platforms have stopped experimenting with generative AI. They now consider it a regular budget line, on par with salaries, cloud services, and energy costs. For context: this is comparable to a large startup's entire infrastructure spending.
Coding costs the most
Almost the entire budget will go to AI agents that work with code: write functions, add unit tests, perform refactoring, integrate services. This is the most expensive and most automatable part of development. Here's Benioff's logic: a junior developer writing routine work 40 hours a week costs roughly $200k a year. An AI agent consuming tokens worth $X a year costs significantly less. Plus, it completes the task 5 times faster.
"This makes everything we build cheaper," —
Mark Benioff on All-In.
Why Anthropic specifically
The choice of provider is not accidental. Salesforce chose Anthropic from all companies creating LLMs because:
- Code quality — Anthropic performs better than others in code accuracy
- Reliability — we need agents that won't break production systems
- Long-term partnership — $300 million is not spent on an unproven player
For Anthropic, this is the first truly large contract deal for enterprise-scale AI coding.
Market trend you'll see
When Salesforce openly announces such figures, the rest of the market watches carefully. Expect Amazon, Google, Microsoft to start publishing similar investments. Expect junior developer salaries to start being calculated with AI agent replacement in mind. This is a signal for engineers: if you write routine code, it's time to learn to work with AI alongside you, not wait for replacement.