Anthropic: agents burn 4x the tokens of chat — 5 ways to cut the bill

The piece is the fourth in a series on what API behavior reveals about AI, and it builds on its own earlier installments: the third explained why agents call models and tools repeatedly, reloading system prompts and conversation history each time, and the second covered auditing. That mechanism is what turns cheap tokens into an expensive bill, because a single task can trigger many calls.

The five fixes are presented as a ranked toolbox, and the article is careful about how much each is worth. Prompt caching is the clearest win, with Anthropic charging a tenth of normal input for cached tokens and OpenAI halving prices automatically; batching halves many services' rates and pairs well with caching. Routing, compression and tool narrowing are more conditional — the RouteLLM numbers (a quarter of requests to frontier models, about 95% of that quality, up to 85% cost cut) come from benchmark conditions, and the article suggests a third of the hoped-for savings is a safer real-world assumption. Tool narrowing is credited with halving prompt tokens and tripling selection accuracy, but compression carries a risk of dropping information.

The article's own caveat is that these measures stack but also erode quality bit by bit, and routing failures are especially quiet because they produce no error, only weaker answers. That is why it ends on measurement: teams can't cut what they can't see, and it points to FinOps-style cost visibility, already common for cloud spend, being applied to tokens. The stakes, then, hinge less on any single technique than on whether organizations can attribute token consumption to a specific team, agent, task and tool — the body notes this is unresolved in many workplaces, and that the next installment moves to implementation and the MCP specification formalized on July 28, 2026.