YC-Bench shows GLM-5 nears Opus

This is a strong reminder that long-horizon agent work is less about one-shot intelligence and more about disciplined state management, memory, and resisting strategy drift.

• –Persistent scratchpad use looks like the real differentiator here; models that kept and rewrote notes consistently outperformed those that didn’t.
• –The benchmark is valuable because it exposes failure modes most evals miss: accepting bad work, looping on stale plans, over-parallelizing, and missing delayed negative feedback.
• –GLM-5’s result is the headline for builders: near-frontier performance at a tiny fraction of the cost changes the economics of production agent pipelines.
• –Kimi-K2.5 standing out on revenue-per-API-dollar reinforces that “best model” and “best model to run at scale” are no longer the same question.
• –Because YC-Bench is open-source and reproducible, it should be a useful stress test for teams building agents that need to stay coherent over long workflows, not just pass short benchmarks.