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.
More from coding & agent
- 818 Open-Source Cybersecurity Skills Empower AI Agents Across 6 Frameworks — tom_doerr · 2026-08-31
- Using Grok bot to handle web scraping tasks without APIs — PTrubey · 2026-08-31
- Why run multiple coding sessions with AI agents? — Jason-Ping · 2026-08-31
- Doug and Hugo Launch Enterprise AI Agent Building Course — hugobowne · 2026-08-31
- From RAG to Agents: Rearranging Search Building Blocks for Better Retrieval — hugobowne · 2026-08-31
- Open source Gambit project uses cross-domain frameworks to aid strategic planning — kidwonder · 2026-08-31