All videos
0:00 / 0:00
ai

Sakana and OpenRouter Say They Beat Fable 5. With Everyone Else's Models.

The AI Automators24 June 2026Watch on YouTube

Part of series

Ep. 3 · Fugu Ultra: Multi-Agent AI

Diepgaande verkenning van Sakana AI's Fugu Ultra als parallel multi-agent systeem dat meerdere modellen simultaan inzet.

View the series

What you'll learn

  • Fugu and Fugu Ultra from Japan's Sakana AI orchestrate a pool of frontier models such as Opus, GPT and Gemini behind a single API, fundamentally unlike a monolithic model.
  • The benchmarks presented are vendor-produced and not independently verified, something the video's maker puts little stock in.
  • Routing (picking the best model per task) and fusing (querying several models in parallel and having a judge merge the answers) are two different architectures with their own cost profiles.
  • Fugu is a fully closed and hosted system without available weights, so as a customer you cannot see which model produced which answer and why.
  • Both vendor lock-in and the original provider lock-in apply: anyone who wants to own the logic should build their own router or fusion layer.

Frequently asked questions

What exactly is Sakana Fugu?
Fugu is a multi-agent orchestration system from Japan's Sakana AI delivered as a single model. Behind one API endpoint, a specialist orchestrator picks from a pool of frontier models such as Opus, GPT and Gemini, delegates subtasks, checks the work and stitches the answer back together. The expensive Fugu Ultra variant throws an entire team at hard tasks.
How does Fugu differ from OpenRouter Fusion?
Fugu is a routing architecture that picks the best model per task and sends a single request there. OpenRouter Fusion instead queries several models at once and has a judge model merge the answers into one final response. Fusion therefore costs more, since you pay for multiple answers plus a verdict each time.
Why does the video's maker put little stock in Sakana's benchmarks?
The benchmarks were produced by Sakana itself and not reproduced by any external party. Moreover, the comparisons with Fable 5 and Mythos Preview are based on Anthropic's published figures, because those models were not publicly accessible for direct head-to-head testing by Sakana.

Topics

mixture-of-modelsmodel-routingensemble-methodsfrontier-modelsllm-architecture

Read next

Sources

What is known about this topic outside the broadcast, and where it says so.

Description from the channel

👉 Access our Starter Apps & AI Architects course in our community https://www.theaiautomators.com/?utm_source=youtube&utm_medium=video&utm_campaign=tutorial&utm_content=fugu 🔗 Sakana Fugu Announcement: https://sakana.ai/fugu/ Trinity (arXiv): https://arxiv.org/pdf/2512.04695 Conductor (arXiv): https://arxiv.org/pdf/2512.04388 🔗 OpenRouter Fusion Announcement: https://openrouter.ai/blog/announcements/fusion-beats-frontier/ Fusion: https://openrouter.ai/fusion A research lab most builders have never heard of just claimed it beats Fable 5. The twist is that Sakana Fugu isn't a model at all. You hit one endpoint, and behind it a trained orchestrator picks from a pool of frontier models, delegates the work, verifies it, and stitches the answer back together. It can even call copies of itself. A week earlier, OpenRouter shipped Fusion: the same bet in a different shape. Stop leaning on one model. Send the prompt to a panel and let a judge merge the answers. Two labs, one idea. So this video answers the question underneath both launches: is a mixture of specialist models actually better than one general-purpose model? #AI #AIAgents #SakanaAI #Fugu #OpenRouter #Fusion #MixtureOfModels #ModelRouting #MixtureOfAgents #LLMRouting #AIArchitecture #AIArchitects #AIBuilder #Claude #Opus #AgentArchitecture