Vals AI: Agentic Tasks Rack Up 10,000x the Environmental Footprint of Simple LLM Queries

Independent benchmarker Vals AI measured carbon emissions, water use and electricity across 16 open-weight models — 14 of them Chinese — running realistic agentic workloads and found lengthy 'thinking' tasks like building a web app carry roughly 10,000x the environmental impact of a one-shot question. Building a single web app on some models is equivalent to powering a home for 2.5 hours; Kimi K3 shows the largest per-task footprint, with DeepSeek V4 Flash only marginally less accurate at a fraction of the cost. Vals engineer Omar Almatov: 'Footprints scale dramatically because these are no longer just single-shot questions.'