Skip to content
AI HOT (Curated Pool)

Agent Workloads Quietly Reshape Inference Economics

智能体工作负载悄然重塑推理经济

SemiAnalysis analyzed 432,000 real coding-agent requests and found a median input length of 96,000 tokens, not 32,000 or 64,000. The post does not disclose the model mix, cost curve, sampling method, or time window.

Why it matters: SemiAnalysis counted 432,000 real coding-agent requests and found a median input of 96,000 tokens, well above the commonly cited 32,000 or 64,000.

Read the original ↗Export Markdown