Cheaper AI inference puts more pressure on the software around the model

The New Stack argues that falling inference prices are shifting value toward the “harness”: software that supplies context, connects tools, routes work and checks results. It points to Zed's agent-oriented code collaboration, OpenRouter's U.S. routing and Anthropic's unified interface, alongside Vercel data showing lower per-token costs. A small private-code benchmark still left every tested coding setup below 40%, suggesting that cheaper calls do not solve integration and verification failures by themselves.