Sakana Ships Fugu Max and Ultra v2 Orchestrators
- Source
- sakana.ai
- Author
- Sakana AI
- Date
Sakana AI released two new systems, Fugu Max and Fugu Ultra v2, that do not compete as single models but as orchestration layers sitting on top of a pool of open weight and specialized models, including NVIDIA's Nemotron. Instead of picking one closed frontier model and sending every request to it, Fugu routes each task to whichever model in its pool handles that kind of work best, mixing cheap models for simple steps with stronger ones for complex, multi step reasoning. Sakana claims the combined system matches or beats Opus 5 on its benchmarks at a fraction of the cost, and because the pool is swappable, the approach avoids locking a product to any single vendor's pricing or availability. This extends a wider pattern this year of orchestration layers competing directly with single frontier models rather than just wrapping them, betting that smart routing across many specialized models can outperform one large generalist on both cost and capability.
Context
Sakana AI's Fugu is an orchestrator: a model that handles a request by deciding which other AI models in a pool to call and how they should work together. Sakana's technical report describes Fugu models as language models "trained to understand user queries and dynamically devise agentic scaffolds" to solve them, and earlier Intel coverage of Fugu describes Sakana offering that system through a single API. The new Fugu Max and Fugu Ultra v2 use that same orchestration design tuned for two different goals, and Sakana says they "are not separate products."
Fugu Max is the cost-focused version. Sakana says it draws on its largest pool yet of open-weights and specialized models, including NVIDIA's Nemotron family, and routes each task to "the leanest model capable of solving" it. It costs $2 per million input and $6 per million output tokens, which Sakana says makes its output pricing 40 to 60 percent lower than Sonnet 5, GPT 5.6 Terra and Kimi K3, while delivering performance "within striking distance of elite models" at two to six times lower cost. Sakana's yardstick here is the cost-performance Pareto frontier: the set of options for which no alternative scores higher without costing more. It reports that Max pushed past that frontier on seven of ten benchmarks and had the best overall score on six.
Fugu Ultra v2 is the quality-focused version. Sakana reports it scored 48.3 on Chartography, a visual reasoning and data interpretation , against 27.3 for Opus 5 and 29.5 for Fable 5, and 74.3 on the software engineering benchmark DeepSWE, where it says Ultra v2 beat models costing three to five times more per token without naming them. Sakana says Ultra v2 reached these results without Fable 5, Fable 5.1 or GPT-6-Astra in its pool of models, and a chart note gives the system's as August 28, 2026. Sakana presents that exclusion as evidence that a swappable pool of open and specialized models "outperforms closed ecosystems" while protecting customers from vendor lock-in and API revocations.
Every figure here comes from Sakana's announcement, which does not say how the competing models' scores were obtained. One benchmark both versions top, SWEFish, is internal and reflects "Sakana AI's own coding challenges." The technical report dates from June 2026 and describes the earlier Fugu and Fugu-Ultra, so it explains the orchestration method but does not test Max or Ultra v2. Both are available through Sakana's OpenAI-compatible API, and existing Fugu users switch with what Sakana calls "a single-line parameter change."
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- knowledge cutoff — The date after which a model saw no training data — everything later has to reach it through search, tools, or the context window.
Checking sign-in…
Loading comments…