GLM-5.2 cut Hermes Agent costs in a week-long setup
Hermes Agent ran on a Hetzner virtual server with , a skill system, scheduled jobs, and more than 60 tools. The setup was used for research, content drafts, management, and git work. The previous model was , which worked well but became expensive because can use many tokens while running for long periods.
GLM-5.2 is an from Z.ai, and it worked in Hermes Agent after changing one config line. There was no fine-tuning and no major prompt rewrite. On OpenRouter, GLM-5.2 cost $0.93 per 1 million input tokens and $3.00 per 1 million output tokens, about 5 to 16 times cheaper than the compared Claude models.
In benchmarks, it scored 74.4% on FrontierSWE, 62.1 on SWE-bench Pro, 77.0 on MCP-Atlas, and 81.0 on . It did not beat Opus 4.8 everywhere, but it stayed competitive, and multi-tool calls worked cleanly.
Key points
- Hermes Agent was running with , scheduled jobs, and more than 60 tools.
- GLM-5.2 was added with one config change, without fine-tuning or a major prompt rewrite.
- OpenRouter pricing was $0.93 per 1 million input tokens and $3.00 per 1 million output tokens.
- The claimed cost gap versus the compared Claude models was about 5 to 16 times cheaper.
- Benchmark results were close enough to make GLM-5.2 a serious cost-saving option for Hermes Agent.