Anthropic's Economic Index: businesses use the API to delegate, not to chat

On September 15, Anthropic published its latest Economic Index, an analysis of how its models are used. It is not a market survey; it describes one vendor's traffic. Even so, the enterprise API findings are among the more concrete public data on how businesses deploy models.

What the report found

Anthropic reports that 77% of enterprise API transcripts show automation patterns, mostly full task delegation, versus 12% for augmentation, where a person works with the model; on its consumer product Claude.ai, the two are nearly even ^[1]. At the task level, 97% of API tasks are automation-dominant, compared with 47% on Claude.ai ^[1].

By content, just under half of API traffic maps to computer and mathematical tasks, more than 8 percentage points above Claude.ai, and office and administrative tasks are about 10% ^[1]. Education and library work is far smaller on the API (3.6%) than on Claude.ai (12.3%) ^[1].

On price, higher-cost tasks generally see more usage, suggesting cost is not the main driver of enterprise adoption ^[1]. After controlling for task characteristics, a 1% cost increase is associated with about a 0.29% drop in usage, so a 10% cost cut would raise usage only about 3% ^[1]. The report calls this a preliminary exploration, and the raw correlation between cost and usage is positive; the negative elasticity appears only after controls ^[1]. Anthropic concludes that model capability and the economic value of automating a task matter more than price ^[1].

How to read it

This is observational data from one vendor's customers: the API sample covers about half of first-party API traffic from August 2025, and coding tasks are just under half of it ^[1]. It says what early API adopters do, not what the average company does, and the report does not measure whether the automation delivered returns. Still, it supports a practical point: production use is shifting toward delegating whole tasks, which puts more weight on testing, checkpoints and review.

Practical implication

If your pilots assume a person reviews every output, check whether the process you actually want is delegation, and design the controls for that case. Choose models on task quality first and price second.