Agent coding input ratios hit 261:1; output token cost becomes irrelevant

xeophon · x · 2026-08-31

Analysis of DeepSWE traces reveals massive input-to-output token ratios in agentic coding: 150:1 for GPT-5.5, 200:1 for K2.7-Code, and as high as 261:1 for Sonnet 5. The author reports a 390:1 ratio for their own Sol usage. This indicates that in agentic workflows, inference costs are dominated by context input, making the cost of output tokens negligible.

Original post →

More from coding & agent

coding & agent channel →