SOMA 把 Astra 长提示词压缩近四成

nordin.eth @nordin_eth原推发布于 2026.09.19

在 X 上打开原推

SOMA’s compression gains are starting to compound. Subnet 114 already cut a ~12.9K-token GPT-6 Astra prompt to ~7.8K tokens, bringing estimated input cost down 39.7%. On GitHub Copilot with DeepSeek V4 Pro, estimated savings have also moved from ~10% at launch to ~15% now. That second part matters most. The compressor isn’t improving because someone shipped a scheduled v2. Miners on subnet 114 are competing to compress agent context better, and those gains feed back into the product. The incentive loop is the upgrade cycle. @SomaSubnet $TAO

这条是引用着另一条写的：x.com/somasubnet/status/2097363100869812645