PointFive Research Finds Cutting AI Tokens Can Actually Increase Costs

New research from PointFive benchmarks nearly 3,000 coding sessions and finds ~80% of AI “bill” spend goes to re-sending cached context rather than generating new answers. The findings challenge the assumption that cached context meaningfully reduces compute cost, implying higher-than-expected inference efficiency costs. The news is more methodological than operational, but it raises cost-efficiency concerns for AI tooling.

Analysis

New research from PointFive benchmarks nearly 3,000 coding sessions and finds ~80% of AI “bill” spend goes to re-sending cached context rather than generating new answers. The findings challenge the assumption that cached context meaningfully reduces compute cost, implying higher-than-expected inference efficiency costs. The news is more methodological than operational, but it raises cost-efficiency concerns for AI tooling.