Unverified50% confidenceBenchmarkExact time
FrugalGPT（Chen et al., 2023）提出基于级联调用的路由框架，先用最廉价模型回答，仅当置信度低于阈值时升级到更昂贵模型，实验表明可在保持 GPT-4 级别准确率的同时削减约 98% 的 API 费用
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
UnverifiedFrugalGPT is a study published by Stanford in 2023 that demonstrated that through cascade strategies, inference costs can be reduced by up to 98% while maintaining GPT-4-level quality.84% similarUnverified斯坦福大学开源的RouteLLM系统实测显示，在保持GPT-4级别输出质量的前提下可将成本降低40%-70%76% similarUnverified斯坦福大学2024年发布的RouteLLM框架研究了三类路由方法，发现在MT-Bench等基准上合理路由策略可在保持95%质量下将GPT-4调用量压缩至约30%，成本节省高达70%76% similarUnverified对低调用量的内部工具型Agent直接调用GPT-4 API总成本可能更低，对高并发低延迟C端产品私有化微调模型长期ROI更优76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/544378API
curl https://kongchang.com/api/v1/knowledge/claims/544378MCP
get_claim(id=544378)