Engineer Reports High Input to Output Token Ratios
Research engineer Florian Brand examined DeepSWE traces for agentic coding tasks.
Florian Brand posted his review of DeepSWE traces. He stated input to output token ratios reached roughly 150 to 1 for GPT-5.5, 200 to 1 for K2.7-Code, 261 to 1 for Sonnet 5, and 390 to 1 in his own Sol usage. DeepSWE itself showed 108 to 1. Brand added that output tokens do not matter in these workloads. The post presents the figures as observations from the traces he reviewed.
Combined views
189
1 post, first seen 3h ago
Engineer Reports High Input to Output Token Ratios
Research engineer Florian Brand examined DeepSWE traces for agentic coding tasks.
Florian Brand posted his review of DeepSWE traces. He stated input to output token ratios reached roughly 150 to 1 for GPT-5.5, 200 to 1 for K2.7-Code, 261 to 1 for Sonnet 5, and 390 to 1 in his own Sol usage. DeepSWE itself showed 108 to 1. Brand added that output tokens do not matter in these workloads. The post presents the figures as observations from the traces he reviewed.
Reactions
1 postFB 
Florian Brand@xeophonlooked at DeepSWE traces and for agentic coding the input:output ratio is ~150:1 (GPT-5.5), 200:1 (K2.7-Code) or even higher, eg 261:1 for Sonnet 5. ccusage reports 390:1 for my own Sol usage (DeepSWE has 108:1) Output tokens don’t matter
- likes: 4
- replies: 0
- bookmarks: 1
- reposts: 0