AI · SEMICONDUCTOR · FACT CHECK
SKD WIREMostly True
AI software company Appier has introduced SMITH, a framework for AI agents that claims to reduce inference token costs by 32 times, with the research accepted at NeurIPS. The claim is based on the company's press materials and published paper.
According to Appier, the SMITH framework enables AI agents to go beyond simply using existing tools: the agents learn to construct their own tools as needed. The company positions this as a step toward more autonomous, cost-efficient agent systems built on large language models (LLMs).
The headline figure — a 32-fold reduction in inference token costs — comes from Appier's own research materials accompanying the NeurIPS-accepted paper. The company presents the token-cost savings as a core result of the SMITH approach, aimed at lowering the expense of running AI agent workflows, where repeated model calls can drive up inference costs.
The research behind SMITH was accepted at NeurIPS, a leading machine learning conference, and Appier has published accompanying materials describing the work. Details of the benchmark conditions underlying the 32x figure are found in the company's paper and press release.
Appier's claim of reducing inference token costs by 32 times with the SMITH framework matches the figures and findings presented in the company's own research and paper, though the results rest on the company's reported measurements — making the overall assessment Mostly True.
