AI Agents Cut Token Bills in Half with Cablese

🔥 What happened A new benchmark reveals that AI models writing in telegraph-style "cablese" cut token costs by up to 49% while other models read the compressed records as well or better than plaintext. The technique works across all tested model families with no training or API changes. 💡 Why it matters For agent-to-agent handoffs and memory stores, this means nearly double the context capacity or half the output bill. But compression rates vary wildly: Gemma hits 40%, Qwen 49% – a measurable per-model property that translates to real money at scale. ⚡ Our take If you're not wiring this into your agent pipeline today, you're literally burning cash. The one exception: mandatory-reasoning models like GPT-5-mini, where compression triples thinking effort and doubles the bill.