Breaking

Sakana AI Unveils Multi-LLM 'Fugu Ultra,' Claims Claude-Level Benchmarks

June 22, 2026 at 22:50 EDT

Foundation Models
AI Agents
Software Dev & Coding

Japanese AI startup Sakana AI on June 22, 2026 released "Sakana Fugu," a system that dynamically routes and orchestrates multiple frontier LLMs, claiming its flagship "Fugu Ultra" matches the performance of Anthropic's top Claude models "Fable" and "Mythos" without the risk of export controls.

June 22, 2026 · Sakana AI

One API, a Swappable Pool of Frontier Models

Sakana AI launches Fugu — a multi-agent orchestrator trained to select, delegate, verify and synthesize across GPT, Claude and Gemini. The top tier Fugu Ultra claims frontier-class performance while routing around single-vendor export-control risk.

73.7

SWE-Bench Pro — Fugu Ultra, ahead of Opus 4.8 (69.2)

Token context window on Fugu Ultra

~500

Beta users who validated the system

SWE-Bench Pro — drawn to scale

Higher is better · score out of 100

73.7

Fugu Ultra

69.2

Opus 4.8

59.0

Fugu

58.6

GPT-5.5

How the orchestrator works

Single API call

→

Fugu selects & delegates

→

Pool: GPT · Claude · Gemini

→

Verify & synthesize

Because the pool is swappable, Fugu can dynamically route around single-vendor restrictions — pitched as an AI-sovereignty play. Fugu Ultra pricing: $5 / $30 per million input/output tokens (≈$10 / $45 above 272k).

Praise from beta use

Code review surfaced 20+ issues where other tools flagged three
Security eval ran recon & XSS/SQLi checks, stayed strictly in scope
High persona stability across long agent sessions

Open questions

Premium pricing while routing to cheaper models — margin concerns
Benchmarks largely self-reported; independent checks advised
Higher latency; trails Fable 5 on certain metrics

Continue reading

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