Frontier Notes / Daily Signal Report
Today
Sakana Fugu, an AI orchestration API routing tasks to multiple frontier models, failed to deliver performance gains over individual models like Claude Opus 4.8 in practical tests while being 5x more expensive and 4.5x slower.
Editor's Notes
This week’s videos reveal a growing focus on AI orchestration (Hermes Agent, Sakana Fugu) as a way to break free from single-model dependency, but hands-on tests show significant cost and speed trade-offs. Meanwhile, an architectural deep dive into GPT and Anthropic’s internal Claude Skills patterns offer grounding in how these models work and how to use them effectively.
Key Takeaways
Hermes Agent allows swapping different AI models for different tasks to optimize cost and performance, breaking the Claude/ChatGPT duopoly. Minimax M3 is highlighted as a particularly cost-effective model that rivals top-tier models at a fraction of the price, using sparse attention for efficiency. The creator demonstrates how to integrate Minimax M3 with Hermes Agent via Telegram for multimodal, web-scraping, and voice-interaction tasks.
GPT (Generative Pre-trained Transformer) is the core architecture behind modern LLMs from labs like OpenAI, Anthropic, and DeepSeek, each tuning it for faster token generation, longer context, better tool calling, and more intelligence. The architecture builds from token embedding (giving tokens internal representation) and positional embedding, through multi-headed attention (Q, K, V vectors for relational communication), to feed-forward networks, layer normalization, and residual connections—all stacked in blocks to predict the next token.
Sakana Fugu Ultra is not a standalone model but an orchestration API that routes tasks to multiple frontier models like Opus, GPT, and Gemini to achieve benchmark results matching Fable and Mythos. In practical tests across 38 tasks, Fugu tied with Claude Opus 4.8 on 36 tasks but was 4.5x slower and 5x more expensive, leading the creator to conclude it isn't worth the cost for his knowledge work, though the orchestration approach is seen as the future of AI efficiency.
Sakana Fugu is an AI routing system that orchestrates multiple frontier models like Opus 4.8, Gemini 3.1 Pro, and GPT55, claiming superior benchmarks. In hands-on tests, it produced functional browser OS, subway scenes, and games, but often fell short of directly using individual models like GPT55, especially for 3D tasks, and incurred higher costs without clear performance gains.
Anthropic employees use Claude Skills in five key ways: categorizing skills into four types (utility, verification, data enrichment, orchestration), leveraging power components like scripts and templates, focusing on verification skills for quality, adding 'gotchas' as living documentation, and tuning triggers for automatic invocation.