Most benchmarks for AI agents measure whether a task got completed and treat however many tool calls, searches, or model upgrades it took to get there as an afterthought. EcoAgent-Bench argues that is backwards for anything running in production, where every action an agent can take, a cheap local lookup, a broad web search, an expensive research tool, escalating to a stronger and pricier model, or looping in a human, has a real cost, and picking the right one for the situation is itself part of the task. The benchmark gives each of its 304 tasks, adapted from existing datasets like GAIA, HotpotQA, and MuSiQue, an explicit budget and tests four specific judgment calls: not escalating when it is not warranted, escalating when local evidence really is insufficient, choosing an appropriately sized model for a given sub-task, and knowing when to just stop because the premise of the question does not hold up. The headline finding is that simple always-escalate strategies can look good on a standard success-rate metric while quietly failing every task where the economical move was to not escalate, which is why the authors also report a separate economic-consistency score that specifically penalizes that one-sided behavior. Under that stricter measure, the tool-using agents they tested scored as low as 3.9 percent to 24 percent on raw task success, and no better than about 7.3 percent on economic consistency, meaning most agents are still either stopping too early when they should dig further, or burning budget on tasks that a cheap lookup would have solved. A further test that gradually tightened the budget found that GPT-5.4's tendency to escalate barely moved at all until the budget crossed a specific threshold, then did not move much further after that, suggesting today's agents are not reasoning about cost so much as reacting to blunt trigger points. For anyone building agent products where API and compute costs actually matter, this is a concrete reminder that whether a task got done and whether it got done economically are different questions, and most current systems are only being evaluated on the first one.