In this comprehensive review, Jack tests two AI models, Sakana Fugu and GLM 5.2, which have been touted as potential challengers to Claude Fable 5. He evaluates their performance, cost, latency, and practical usability within the Hermes agent system. The findings reveal that while Sakana Fugu offers intelligent routing, it adds latency, and GLM 5.2 impresses with speed and cost-efficiency but is not a definitive Fable 5 killer yet.
In the rapidly evolving world of AI language models, keeping up with the latest and greatest can be exhausting. Recently, two models have been making waves as potential "Fable 5 killers": Sakana Fugu and GLM 5.2. In this article, I, Jack, share my in-depth testing of these models within the Hermes agent system to determine if they truly outperform Claude Fable 5 and which models are best suited for use today.
Sakana Fugu is not a traditional model but an intelligent router. It acts as a single API that orchestrates a pool of Frontier models, deciding which model to call based on the type of question asked. For example, it might route Gemini tasks to one model, Opus 4.8 to another, and so forth. This multi-agent system is designed to compete directly with Fable 5 by leveraging the strengths of multiple models.
GLM 5.2 is an open-weight model from China, known for its cost-effectiveness and performance. It reportedly matches or exceeds Opus 4.8's capabilities at roughly one-sixth of the cost. It supports a large context window (up to 1 million tokens) and can be run locally or integrated into Hermes agent.
If either Sakana Fugu or GLM 5.2 can outperform Fable 5, it means Hermes agent users can access a more capable AI brain for less money. This has significant implications for AI startups and developers seeking efficient, powerful AI solutions.
To use Sakana Fugu with Hermes agent, you need to create an account on the Sakana website, generate an API key, and securely provide it to Hermes via a terminal command. Once connected, you can instruct Hermes to use the Sakana model for tasks.
GLM 5.2 can be connected directly via its API key or through Open Router, which simplifies integration. Documentation and guides are available to assist with setup, including connecting GLM 5.2 to Claude Code.
I tested both models with simple prompts to verify functionality:
Both models performed well on these basic tasks, with GLM 5.2 correctly counting letters and choosing to drive, which aligns with practical reasoning.
I tasked each model (GLM 5.2, Sakana Fugu, and Opus 4.8) to retrieve the subject line of my last email from Outlook using the Zapier MCP tool. Sakana Fugu won this test, completing the task with the fewest tokens and correct output. GLM 5.2 initially failed but succeeded on the third attempt. Latency was lowest for GLM 5.2 and Opus 4.8, with Sakana Fugu being slower due to its routing overhead.
Each model was asked to create a simple, visually engaging one-page website for a sparkling water company with three sections and links to test the models locally.
Despite the simplicity of the prompt, GLM 5.2 showed superior efficiency and output quality.
The models were challenged to improve the memory tab of the Hermes and Claude Code agentic operating system using Graphify for codebase understanding.
GLM 5.2's contribution was the most visually appealing and functional.
When choosing models for Hermes, consider four key factors:
My recommended approach is to use Hermes as a black box that routes tasks to the most optimal model based on the question type. This can be done dynamically via voice commands or by specifying models for particular skills within Hermes.
For example, use GLM 5.2 and Opus 4.8 for complex "big brain" tasks, have them debate or collaborate, and use cheaper models like DeepSeek for high-scale, less critical tasks.
While Sakana Fugu and GLM 5.2 bring exciting innovations to the AI model landscape, neither is a definitive Fable 5 killer yet. Sakana Fugu's intelligent routing is promising but costly in latency and tokens. GLM 5.2 stands out for its cost-efficiency and surprisingly strong performance, making it a valuable addition to the Hermes agent ecosystem.
Understanding the strengths and weaknesses of each model and leveraging Hermes agent's flexible routing capabilities will enable users to maximize AI performance while managing costs and privacy.
Stay tuned for further updates as these models evolve and new challengers emerge in the AI space.
