Break-Even Math for Self-Hosted GLM-5.1: When Claude API Rates Win
The reported 10x saving from self-hosting GLM-5.1 instead of paying Claude API rates is a utilization claim, not a hardware claim. This video walks the arithmetic: an 8-GPU node rented at $2.50/GPU-hour costs $20/node-hour, and that hourly price has to be divided by the utilization you actually sustain. At 100% utilization the coding workload (40k input, 1.5k output, 80% cache hit) comes out about 9.4x cheaper than Sonnet list prices. Break-even against those list rates is 10.6% utilization — 7.0% on owned hardware at $13.15/node-hour, and above 20% if the API side uses Batch pricing. At 5% utilization the API is ahead by roughly 2.1x. Traffic shape moves the number: input-heavy retrieval breaks even near 6.7%, while agentic loops with long outputs need 14.9%. Decode has about 5.4x headroom over the API output price, prefill about 13.6x — long outputs are the worst case for self-hosting, not the best. The post also lists the seven numbers you'd need to audit any '10x cheaper' claim, including prompt cache hit rate and sustained utilization across nights and weekends. Throughput figures are stated assumptions, not benchmarks of GLM-5.1; the arithmetic is exact. Full write-up: https://hackyjs.com/posts/break-even-math-for-self-hosted-glm-51-gpu-utilization-ops-overhead-and-claude-api-rates