Breaking

Sakana AI Launches "Fugu," Linking Multiple Models, Stresses Export-Control Workaround

June 22, 2026 at 08:14 EDT

AI Agents
Foundation Models
Sales & Business

Tokyo-based Sakana AI began full availability of "Sakana Fugu," a multi-agent orchestration system that dynamically coordinates several frontier foundation models, around June 22, 2026. The higher-tier "Fugu Ultra" claims performance rivaling Anthropic's Claude Fable 5 and Mythos Preview on some benchmarks, while emphasizing that it is less exposed to U.S. export-control risk.

June 22, 2026 · Sakana AI

A 7B Orchestrator, Not a Bigger Model: Sakana Fugu Goes Commercial

Fugu coordinates multiple frontier models through one API. Sakana claims its top-tier Fugu Ultra matches Anthropic's Fable and Mythos — "without the risk of export controls."

A small Conductor model orchestrates far larger frontier models

Variants — latency-first Mini and deep-reasoning Ultra

20+

Issues surfaced in one code review vs. ~3 from conventional tools

SWEPro Benchmark — score to scale

Higher is better. Fugu Ultra leads the field on coding tasks.

54.2

Fugu Ultra

53.4*

Opus 4.6

51.3

Fugu Mini

51.2

GPT 5.4

48.4

Gemini 3.1

* Opus SWEPro uses the vendor's own reported value; some figures are self-reported on specific scaffolds.

How the orchestrator works

Conductor

Selects & delegates

→

Agent Pool

Frontier models run

→

Verify & Synthesize

Recursive self-calls

The Pitch

"Fable/Mythos-class" capability without export-control risk
No single-vendor dependence
Strong on complex workflows, research, strategy, security

The Doubts

Ultra latency of tens of seconds to minutes
Frontier rates while routing to cheaper models
No transparency on which model answered; self-reported benchmarks

Fugu Ultra pricing

$5 / M input tokens

$30 / M output tokens

Standard plan $20/mo + usage

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.