Sakana AI launches Fugu Max and Ultra v2 orchestration models

Sending a trillion-parameter giant to run a basic database lookup is not peak intelligence. It is just expensive.

Sakana AI released Fugu Max and Fugu Ultra v2. Instead of building monolithic models from scratch, Sakana is doubling down on orchestration, routing tasks across large pools of open and specialized models to balance cost and capability.

Split missions on the same engine

Both systems share the same underlying routing architecture, but they are tuned for entirely different missions.

Fugu Max focuses on squeezing costs without dropping output quality. Priced at $2 per million input tokens and $6 per million output tokens, its output pricing runs 40% to 60% cheaper than Sonnet 5, GPT 5.6 Terra, and Kimi K3. By pulling in open models like NVIDIA's Nemotron family, it topped six benchmarks, including Terminal Bench 2.1, GPQAD, and Sakana's own SWEFish coding benchmark.

Fugu Ultra v2, on the other hand, aims squarely at peak performance on multi-step, agentic tasks. It captured top scores on five of eight benchmarks, logging a 74.3 on DeepSWE and hitting 48.3 on Chartography, easily outpacing Opus 5 at 27.3 and Fable 5 at 29.5.

Crucially, Sakana noted that Fugu Ultra v2 pulled off these scores without including proprietary heavyweights like Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool. Both models drop in seamlessly via standard OpenAI-compatible endpoints with a single parameter change.

Why it matters

We think Sakana is hitting the most pragmatic nerve in enterprise AI right now. Brute-forcing bigger foundation models delivers diminishing returns when real-world production budgets are suffocating under token costs. If an orchestration layer can deliver frontier-level results while cutting output pricing in half, developers have every reason to take notice.

There is also a clear sovereignty play here. Relying on a single closed provider leaves teams vulnerable to sudden price spikes, API policy shifts, or geopolitical blocks. By building an abstraction layer over swappable open models, Sakana offers a compelling escape hatch from vendor lock-in.

The real test will be latency and router reliability under messy, real-world traffic. Multi-agent systems can easily get bogged down if the meta-router misjudges complexity or adds round-trip overhead. Still, this looks like a smart pivot away from monolithic arms races and toward genuine operational efficiency.