Edgee AI Launches Compressor V2, Cutting LLM Agent Costs by Up to 50%

Summary

Edgee AI has released Compressor V2, an upgraded compression layer for its AI gateway that achieves approximately 50% cost reduction for LLM-based coding agents. The new version employs three orthogonal compression strategies—Brevity, Tool Surface Reduction (TSR), and Tool Result Trimming—each targeting different sources of token bloat in long-running agent sessions. This advancement addresses a critical economic constraint: a single SWE-bench coding task can consume 1–10 million tokens across 30–100 API turns, making cost optimization essential for AI products at scale.

Compressor V2 improves upon the V1 version, which relied solely on tool result trimming and delivered ~10% cost savings. The new multi-layered approach is strategically designed to preserve Anthropic's prefix cache benefits—critical for long sessions where system prompts and tool catalogs are reused. By targeting only output tokens and suffix content rather than the cached prefix, the compression strategies maintain cache amortization while concentrating savings in the most expensive token class. Each strategy is independently configurable per API key, allowing customers to tailor the compression combination to their specific workloads.

The results have been rigorously validated using paired sign testing, bootstrap confidence intervals, and within-task coefficient of variation analysis on SWE-bench Lite, employing statistically sound methodology to account for the enormous variance in task difficulty (from 100k to 12M-token tasks).

• Each compression strategy is independently configurable, allowing customers to compose the combination that matches their specific workload requirements