Sakana Fugu Beats Claude Fable 5 with Multi‑Model Orchestration

Sakana Fugu: Multi‑Agent Orchestrator Unveiled

In a bold move to sidestep the sudden withdrawal of Anthropic’s Claude Fable 5, Sakana Fugu offers developers a single OpenAI‑compatible endpoint that dynamically assembles a pool of specialist models.

The system treats a complex request like a construction project: it fragments the job, assigns each fragment to the most capable model, validates the output, then fuses the results into a coherent answer.

How the Orchestration Works

Powered by Sakana’s internal research papers “TRINITY” and “Conductor,” the orchestrator itself is an LLM that can even call copies of itself recursively. It decides in real time which provider—Google, DeepSeek, or emerging open‑source engines—should handle each sub‑task, keeping the selection algorithm proprietary.

Benchmarks show Fugu Ultra achieving ↑ 93.2 on the LiveCodeBench coding suite, edging out the now‑unavailable Claude Fable 5 (89.8). On the GPQA‑D science quiz it posted 95.5, surpassing Anthropic’s preview model (94.6).

“Orchestration is the next frontier beyond raw model size,” said Sakana CEO David Ha on X.

Must Read Intel Review our latest briefing on this sector

Related Intel: AI Context Tier Emerges as New Bottleneck in Inference Workloads

Pricing follows a pay‑as‑you‑go model; the Ultra tier costs $5 per million input tokens and $30 per million output tokens, a figure that translates to a ↓ 30% cost advantage versus single‑vendor APIs for comparable workloads.

Enterprises can blacklist specific providers to meet data‑privacy mandates, and a free second month is offered to early adopters who sign up before July 31 2026.

Analysts note that while the orchestrator shines on multi‑step tasks, monolithic frontier models still dominate pure reasoning benchmarks when unrestricted access is guaranteed. For more on the export‑control backdrop, see Reuters and Bloomberg.

Intel provided by Julian Reed (Consumer Electronics Expert).