Show HN: 2x-4x cheaper GLM 5.3 for coding and research

Same model, same prompt — multiple times the tokens per second, with near-zero rate limits. Inference built for long-running headless agents, so the runs that used to queue now finish on time.