Anthropic’s new beta header trims the token cost of tool calls by about 14 percent for Claude 3.7 Sonnet, giving AI agents a modest reduction in their monthly bills and a small latency edge.

Why tool use eats tokens

Every turn an agent takes with Claude involves three token-driven steps:

  • Tool definitions – the names and JSON schemas you send as part of the prompt.
  • Tool calls – the structured output the model generates when it decides to invoke a tool.
  • Tool results – the data your code returns to the model.

The first and third steps are input tokens; the second step is output tokens. Because Claude charges more for output than for input, cutting output tokens can lower costs even if the overall token count stays similar.

The beta header

Anthropic announced a beta feature named token-efficient tool use. Adding the HTTP header 

anthropic-beta: token-efficient-tools-2025-02-19

activates the optimization, but only when the request targets Claude 3.7 Sonnet. If the header is sent to any other model the request proceeds unchanged; the header is ignored without error.

The optimization works by compressing the model’s tool-call output, shaving roughly 14 percent off the output token count for that step.

How much you actually save

Output tokens are pricier than input tokens, so a 14 percent cut on that slice does not translate into a proportional drop in the total bill. In a typical four-step agent loop, tool definitions often dominate input costs, sometimes exceeding half of the tokens used. Under those conditions the 14 percent saving on output tokens typically yields only about a 3 percent reduction in the overall charge.

The headline-saving figure therefore feels modest, but the change comes at zero extra cost and introduces a slight latency boost: fewer output tokens mean the model finishes generation a fraction quicker.

Bigger savings require different tactics

If the goal is to meaningfully curb expenses, the header alone won’t do it. Three practical levers deliver larger gains:

  • Prompt caching – store the tool definitions once and reuse the cached version on subsequent calls. Cached reads are billed at a lower rate than fresh input tokens.
  • Trim the tool set – only include tools that are essential for the current task. Every extra tool adds its definition to every request, inflating input costs.
  • Slim down tool results – return only the fields the model actually needs. Large API responses that are fed back into the conversation inflate both input tokens and the conversation history.

Applying these practices can shave a noticeable chunk off the total spend, far beyond the 3 percent you’d see from the beta alone.

What to watch

Anthropic has not indicated when—or if—the token-efficient mode will graduate from beta to a default feature. Developers should keep an eye on future announcements that might broaden support beyond Claude 3.7 Sonnet or introduce deeper token-compression techniques. Monitoring the per-request token breakdown will also help quantify the real-world impact as usage patterns evolve.

Bottom line

The beta header is a free, low-effort tweak that trims tool-call output tokens by roughly 14 percent, delivering a modest bill dip and a tiny speed bump. For anyone already wrestling with high agent costs, the header is a helpful add-on, but real savings will come from caching prompts, limiting tool exposure, and returning leaner results.

Source: https://dev.to/multigrid/token-efficient-tool-use-in-the-claude-api-3j0l