Writer's AI harness cuts token spend nearly 40% — without sacrificing accuracy

This shifts the bargaining power in enterprise AI away from raw model access and toward whoever controls the workflow layer. The near-term losers are the “thin wrapper” vendors whose economics depend on customer ignorance of unit costs; if buyers start measuring task-level spend, pricing power compresses fast. The longer-term winners are the platforms that own identity, policy, retrieval, and audit rails — think MSFT, GOOGL, AMZN, SNOW, ESTC, and, to a lesser degree, NOW/PLTR — because harness optimization makes control-plane integration more valuable than model novelty.

The second-order effect is that efficiency can be deflationary at the invoice level but inflationary at the adoption level. If the best-controlled deployments can cut cost per task while holding quality, enterprises will likely reallocate budget from brute-force model spend into more workflows, which supports aggregate consumption but penalizes suppliers that only sell tokens. That argues for a barbell: infrastructure names with pricing discipline and distribution on one side, and short-duration, valuation-sensitive AI application names with weak software moats on the other.

The contrarian miss is that this is not automatically bearish for frontier models. The study implies that the most capable models are the only ones reliable enough to exploit the cheaper orchestration regime, so the top end may capture a larger share of enterprise spend even as waste falls. The real falsifier is not lower token usage; it is whether enterprise AI revenue fails to re-accelerate over the next 2-3 quarters despite improving unit economics. If that happens, the “efficiency unlock” thesis was overestimated and the market should fade the premium attached to AI software rollups.