Atlas

Models

← All models

Fugu Ultra v1.0

Sakana AI · 2026-06-22

Fugu Ultra v1.0 is the initial performance-oriented Sakana Fugu orchestrator released through Sakana AI’s API, now identified as `fugu-ultra-v1.0` and formerly as `fugu-ultra-20260615`. Starting from a pretrained language-model checkpoint, it was trained with GRPO to design query-adaptive workflows of up to five steps over a diverse pool of frontier LLM agents; the hosted model supports text and image input, a 1,000,000-token context window, and configurable high or xhigh reasoning effort.

Benchmark scores

BenchmarkScore
AA-LCR73.3
BuseyBench SVG5.8
Capability162.8
CharXiv Reasoning86.6
Eval Connections100.0
Eval Connections — Guess Accuracy82.5
Eval Connections One-Shot Group Accuracy100.0
Eval Connections One-Shot Puzzle Solves100.0
Eval Connections One-Shot Trap Bonus60.0
Eval Connections One-Shot with Traps84.0
GPQA Diamond95.5
Humanity's Last Exam50.0
KernelBench Hard Normalized (RTX PRO 6000)0.0
LiveCodeBench Pro - 2025 Q2 pass@190.8
LiveCodeBench v6 - full 1,055 problems93.2
MRCR v2 (8-needle) - 128K average93.6
ObviousBench97.9
SciCode (Fugu background, patched libraries)58.7
SWE-Bench Pro (Public Dataset)73.7
Terminal-Bench 2.182.1
τ³ Banking - pass@420.6
Loading Atlas data…