NVIDIA's Hack Halves Coding Agent Costs

🔥 What happened NVIDIA, NTU and MIT dropped SoL-Pi: four efficiency tricks for the open-source Pi coding agent. An AI discovered them itself across 3,000+ auto-research runs. On EdgeBench, token traffic drops by up to 49 percent. 💡 Why it matters API costs fall roughly 33 percent while keeping 94 percent of Pi's score. Instead of faster kernels, this optimizes the harness — the layer handling tool calls and context. That's where most tokens actually burn. ⚡ Our take Running coding agents in 2026 without harness tuning is burning money. The four mechanisms are unglamorous, which is exactly why they're instantly copyable.