AI agents get better mostly by spending more tokens — so cheaper tokens just mean bigger bills, not smaller ones.

This explores why agentic AI's token bills keep growing even as each token gets cheaper. The corpus doesn't track price declines directly, so it can't confirm the race between volume and price. It does explain the volume side well, and the explanation is surprising. Agents use more tokens partly because spending more tokens is how they get better. Anthropic's internal evaluations found that token usage alone explains about 80 percent of the performance differences in multi-agent research systems Does token spending drive multi-agent research performance?. When quality tracks spending that closely, every price cut becomes room to spend more. Builders usually use the savings to run more agents, longer loops and deeper searches, not to cut the bill How does test-time scaling work at the agent level?.

The second reason is waste. Much of what agents consume isn't useful work. When researchers automatically tuned agent harnesses (the scaffolding that manages an agent's actions, memory and tool results) across many environments, they found four mechanisms that cut token traffic by 44.7 to 49.0 percent with comparable performance Can agent harnesses be automatically optimized across many environments?. Microsoft's ACON method cut peak token usage 26 to 54 percent just by rewriting the instructions that tell an agent how to compress its own context Can prompt optimization cut agent token usage without retraining?. Default agent setups, in other words, leave about half their tokens on the table. Another study sharpens the point: agent performance tracks 'effective feedback' (informative, non-redundant signal the agent actually uses) far better than raw token count Does raw token spending actually predict agent performance?. Much of the volume growth is redundant re-reading and repeated context.

This suggests counting tokens may be the wrong measure. A 115-day case study of a persistent agent found that 82.9 percent of its tokens were cache reads, meaning the same context was reused rather than generated fresh. The authors argue the useful unit of cost is the finished artifact, not the token Do persistent agents really cost less per token?. Work on compact models makes the same point from the other side: cost and latency pile up over a whole multi-step episode. A smaller model trained to follow through reliably can beat a larger one on total cost, even if it isn't cheaper per token Does model efficiency matter more than peak capability for real work?.

There's also a bigger-picture reading. One theoretical note argues AI marks a shift from the 'age of the commodity' to the 'age of the token.' Value is made through tokens generated in context at the moment of use, not through identical mass-produced goods, and that dynamic tends toward inflationary devaluation Is AI fundamentally changing how value gets produced?. Cheap, endlessly producible tokens invite more production. Payment data shows this becoming routine: on the x402 agent-payment rails, the median time from funding an agent's wallet to its first spend fell from 73 days to 18 minutes Are AI agent wallets becoming genuinely operational instead of experimental?.

The takeaway you might not expect: the growth in token volume is partly a choice and partly avoidable. The choice is to spend tokens because spending buys performance. The avoidable part is harness inefficiency that roughly doubles consumption. The gains available from smarter context handling appear to be larger than a typical round of price cuts.