Hermes Agent: GLM 5.2 vs Sakana Fugu vs Opus 4.8

The speaker compares Sakana Fugu, GLM 5.2 and Claude Opus 4.8 inside Hermes Agent. The tests cover retrieving an Outlook email subject through Zapier MCP, building a sparkling water website and improving a memory tab in his custom Hermes interface codebase. These are the speaker's practical trials, rather than a general benchmark.

Fugu is described as a router that selects from a hidden pool of models behind one API. The speaker introduces GLM 5.2 as an open-weight model that users can run themselves, though this walkthrough connects through APIs rather than demonstrating a local LLM deployment. Setup covers creating a Sakana API key, entering credentials through a terminal prompt and accessing GLM through OpenRouter or a direct key.

The speaker reports that Fugu won the tool-calling test, while GLM needed three attempts. For the website task, the speaker preferred GLM after requesting a contact form and noted that the prompts lacked website-specific skills. In the codebase task, all three finished; Opus was fastest, while the speaker preferred GLM's memory health bar. Fugu's longer waits and token use weakened its appeal in these trials.

The closing recommendation is to assign models to recurring tasks within the AI agent, weighing output quality, speed, privacy and cost. The speaker demonstrates explicit model choices in skills through his custom interface and favors that approach over relying on Fugu's automatic routing.