Agents

How do AI Agents spend your money? Most teams treat agent token costs as a rounding error even though the data says they shouldn't. New pape…

How do AI Agents spend your money? Most teams treat agent token costs as a rounding error even though the data says they shouldn't. New paper presents the first systematic study of how agents actually

DGX agentx-post
agentsdair-ai--x

How do AI Agents spend your money? Most teams treat agent token costs as a rounding error even though the data says they shouldn't. New paper presents the first systematic study of how agents actually spend money on coding tasks. They ran 8 frontier LLMs on SWE-bench Verified and tracked every token. The findings: - Agentic coding burns ~1000x more tokens than chat or code reasoning, mostly on the input side. - Identical tasks vary up to 30x in token usage across runs. - Higher token spend doesn't buy accuracy. Performance peaks at intermediate cost and saturates. - Models are bad at predicting their own token usage. Self-forecast correlation tops out at 0.39. - Some models burn 1.5M more tokens than others on the same task with no quality gain. If you're shipping agents, runtime cost is high-variance, weakly tied to quality, and unpredictable even by the model itself. That changes how teams budget, route between models, and decide when to kill a run. Paper: https://arxiv.org/abs/2604.22750 Learn to build effective AI agents in our academy: https://academy.dair.ai/

Related

Source: DAIR.AI (X) | 2026-04-27

Loading related sources…