SOMA Claims 39.9% Cheaper Input via Context Compression on the Same Prompt

const_reborn · x · 2026-08-17

The SOMA team tested their SOMARIZER context compressor on a 2D HTML/CSS animation prompt (against ChatGPT 5.5 Pro): the original ran 8,680 tokens / 31,364 characters, an estimated $0.5729 input cost at $10/MTok.

After compression at a 0.60 ratio, it dropped to 5,221 tokens / 17,590 characters and $0.3447 — 43.9% fewer characters and 39.9% lower input cost. They claim compression removes not just noise but redundant and contradictory instructions, leaving a tighter spec for the model. Note: vendor-run benchmark, independently unverified.

Original post →

More from Apps

Apps channel →