Sakana AI Launches Fugu Max and Fugu Ultra v2, Betting on Multi-Agent Orchestration

On Sept 11, Sakana AI released two new versions of its multi-agent orchestration system: cost-efficient Fugu Max and flagship Fugu Ultra v2. The company's core thesis is that the real frontier is the capability × cost Pareto front, not ever-larger single models—dynamic orchestration turns a static model menu into a schedulable system. The next day Fugu Ultra v2 went live on OpenRouter, and Nikkei reported it again reached world-class performance by combining multiple AI models. All performance and cost claims remain officially sourced, without independent verification.

Confirmed

• Both products share the same multi-model orchestration architecture with different positioning: Fugu Max targets cost-efficiency at $2 input / $6 output per million tokens, with Sakana claiming performance near top-tier models at 2–6x lower cost; Fugu Ultra v2 is the flagship for complex multi-step tasks and no longer requires the old scheduling form (per hardmaru).
• According to Sakana AI, Fugu Max uses a "model-pool orchestration" architecture scheduling a mix of NVIDIA Nemotron and open/closed models, with replaceable multi-model routing for resilience; the company calls model resilience "part of the architecture, not a backup feature" and says it orchestrates its largest open-weight model pool to date.
• On Sept 12 the company announced Fugu Ultra v2 on OpenRouter: not a single model but a trained multi-agent orchestration system in which a language model routes tasks to a fixed pool of open and specialized models, and can recursively call its own instances.
• The accompanying blog reports benchmark results: Fugu Ultra v2 achieved best or tied-best on 5 of 8 benchmarks, including DeepSWE 74.3 and Chartography 48.3; per Sakana AI (relayed by hardmaru), it outscored GPT-6 Astra and Claude series on some coding benchmarks like DeepSWE.
• VoidAsuka summarized the philosophy: the industry has spent a decade optimizing on a single bigger-and-pricier axis, whereas intelligence is fundamentally collective and distributed—the strongest AI is a collaborating model ecosystem, not an isolated giant.
• NVIDIA AI also covered the launch, calling Fugu the next evolution of multi-agent orchestration systems and noting the industry still treats the capability-cost frontier as a static menu of isolated models.
• On Sept 12 the company published full release notes, reiterating that the frontier is a two-dimensional capability × cost problem and the future belongs to orchestration systems that know which machine to invoke to solve a task at lowest cost.
• According to Nikkei on Sept 12 (relayed by Sakana AI's account), Fugu Ultra v2 again reached world-class performance through multi-model coordination, including coding, and the report framed it as a representative Japanese corporate achievement.

Unconfirmed

• Claims such as "beats GPT-6 Astra," "2–6x cheaper," "best on 5 of 8 benchmarks," and "world's highest performance" come from Sakana AI, reposts, and Nikkei coverage; no third-party independent evaluation has yet corroborated them.

Why it matters

• If orchestration truly approaches flagship performance at much lower cost, it would upend the stack-parameters race, give budget-sensitive developers a new option, and pressure pricing of closed-source flagship models.
• Fugu Ultra v2's OpenRouter availability lowers the barrier for developers to test the claims; if the multi-model routing's resilience holds, it also offers a new engineering pattern for production systems dependent on a single model API.
• Known for low-cost innovation, Sakana AI is shifting competition from single models to system orchestration, drawing NVIDIA's attention and Nikkei coverage as a flagship Japanese AI result, signaling traction both industrially and in public opinion.

2026-09-11 ~ 2026-09-12 · 16 related posts

• [source] Sakana AI launches Fugu Max and Fugu Ultra v2, matching elite models at 2-6x lower cost — SakanaAILabs · 2026-09-11
• Sakana AI's Fugu Ultra v2 Claims to Beat GPT-6 Astra on Some Benchmarks; Fugu Max Cuts Output Price 40-60% — hardmaru · 2026-09-11
• Sakana releases Fugu Max & Fugu Ultra v2, betting on orchestrated model ecosystems over monoliths — VoidAsuka · 2026-09-11
• Sakana's Fugu Max routes a mixed model pool at $2/$6 per 1M tokens, claims tier-best benchmark score — SakanaAILabs · 2026-09-11
• Sakana AI Ships Fugu Ultra v2, Multi-Model Orchestration Returns to World-Class Performance — SakanaAILabs · 2026-09-12
• [source] Sakana AI's Fugu Ultra v2 multi-agent orchestration engine lands on OpenRouter — SakanaAILabs · 2026-09-12

10 near-duplicate retellings: hardmaru · SakanaAILabs · SakanaAILabs · hardmaru · hardmaru · NVIDIAAI · graceisford · SakanaAILabs · SakanaAILabs · SakanaAILabs