Skip to content
NewsModel release

Sakana AI ships Fugu Max and Fugu Ultra v2, beating frontier benchmarks without frontier models

· by Pondero Newsdesk

The short version

Japan's Sakana AI released two orchestration models on September 11: Fugu Max at $2 per million input tokens and Fugu Ultra v2, which scored 48.3 on Chartography visual reasoning versus 27.3 for Opus 5, without Fable 5 or GPT-6-Astra in its pool.

Sakana AI ships Fugu Max and Fugu Ultra v2, beating frontier benchmarks without frontier models

Sakana AI's Fugu Ultra v2 scored 48.3 on Chartography, a visual reasoning and data-interpretation benchmark, against 27.3 for Anthropic's Opus 5 and 29.5 for Fable 5. It achieved that gap without including either closed model in its orchestration pool. The Tokyo-based lab released Fugu Ultra v2 and a new cost-optimized sibling, Fugu Max, on September 11, 2026, pushing an architecture built around learned orchestration rather than a single monolithic model.

What Fugu is

Fugu is not a foundation model. Per Sakana AI's announcement, it is a language model trained to route tasks across a fixed pool of open-weight and specialized models, and to recursively call instances of itself. The orchestrator decides which sub-model to dispatch for each task, then assembles the outputs. The design is intentionally vendor-agnostic: because the underlying pool is swappable, the system can replace any model that becomes unavailable, degrades, or is restricted by its provider.

The architecture has iterated publicly since April 2026. Sakana released a beta that month to demonstrate orchestration as a unified foundation-model interface. General availability and Fugu Ultra v1 followed in June. July added Fugu-Cyber and a Claude Code interface. August brought NVIDIA Nemotron model integration via a new NVIDIA partnership and a consumer launch inside Sakana Chat. September 11 is the first release shipping two distinct pricing tiers simultaneously.

Fugu Max: cost-first tier

Fugu Max expands the pool of models Sakana can orchestrate, adding the NVIDIA Nemotron family through the company's NVIDIA collaboration, per Sakana AI. The orchestrator routes each task to the lightest capable model in the pool.

Pricing is $2 per million input tokens and $6 per million output tokens. Output pricing sits 40-60% below Sonnet 5, GPT 5.6 Terra, and Kimi K3, per Sakana AI. On benchmark comparisons within a similar price range, Fugu Max achieved the best overall score on six of ten evaluated: Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. It expanded the cost-performance Pareto frontier on seven of ten benchmarks.

The performance claim is bounded by peer pricing. Sakana compared Fugu Max against models "in a similar price range," not against unconstrained frontier models. That scope is appropriate given the pricing tier but matters when evaluating the benchmark headlines.

Fugu Ultra v2: capability-first tier

Where Fugu Max optimizes cost, Fugu Ultra v2 optimizes for peak output quality on complex, multi-step tasks. Its training cutoff is August 28, 2026.

Key benchmark results, per Sakana AI:

  • Chartography (visual reasoning and data interpretation): 48.3 versus Opus 5 at 27.3 and Fable 5 at 29.5
  • DeepSWE (real-world software engineering): 74.3, outperforming models that cost three to five times more per token
  • Best or joint-best on five of eight evaluated benchmarks: GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon
  • Top-two placement on seven of eight benchmarks

Fugu Ultra v2 generated those results without Fable 5, Fable 5.1, or GPT-6-Astra in its model pool, per Sakana AI.

Pricing on OpenRouter is $5 per million input tokens and $30 per million output tokens, per OpenRouter's listing. Prompts above 272K tokens are billed at a higher rate; Sakana has not published the specific figure for that tier. Orchestration tokens consumed internally are billed as standard input/output tokens. The context window is 1 million tokens.

Both models are available immediately via an OpenAI-compatible API. Teams already running Fugu can switch to Max or Ultra v2 with a single parameter change and no migration work required, per Sakana AI.

Why it matters

Two dimensions here are relevant to operators choosing between orchestration systems and direct frontier API calls.

The architecture creates a specific kind of resilience. A direct API call fails the way that provider fails. An orchestration system built on a swappable pool degrades more gracefully when one model goes offline, gets rate-limited, or changes its terms. Sakana describes this as "supply chain resilience" and frames it as a production benefit rather than a fallback. For teams running agentic workloads at scale, the failure mode of underlying model infrastructure is increasingly a design choice, not merely an operational detail to manage after an outage.

The benchmark gap on Chartography is harder to dismiss than a marginal margin. A score of 48.3 against 27.3 is roughly 77% higher on that task type. DeepSWE at 74.3 is the same story on code. Benchmarks measure narrow domains, and Operators running tasks outside visual reasoning and software engineering cannot assume equivalent outperformance. But for AI coding pipelines and document-analysis workflows, both of which are primary enterprise agentic use cases in 2026, those two benchmark domains land squarely in scope.

The cost calculation is more nuanced. Fugu Ultra v2 at $5/$30 per million tokens is not dramatically cheaper than frontier alternatives on output-heavy workloads. The real cost case is Fugu Max at $2/$6, where the 40-60% output pricing discount against Sonnet 5 and GPT 5.6 Terra applies to every token produced. For operators running high-volume tasks that do not require peak-quality reasoning, that delta compounds quickly.

Context and reactions

Sakana AI was founded in 2023 by former Google Brain researchers and is headquartered in Tokyo. Fugu represents a significant pivot from the lab's earlier work on evolutionary and neural architecture search approaches. The NVIDIA partnership, announced in August 2026, gave Fugu Max access to Nemotron open models that would otherwise be impractical to integrate at orchestration scale.

Anysphere's Cursor Projects, a multi-agent cloud environment, shipped in the same week with positioning that overlaps on the "pools of agents for complex tasks" framing. The two products target different layers: Fugu is an API-level orchestration service; Cursor Projects is an IDE-integrated workflow builder. Procurement conversations at enterprise accounts evaluating long-running agentic workloads will likely encounter both.

What to watch next

Independent benchmark labs reproducing Fugu Ultra v2's Chartography and DeepSWE scores against the same frontier models are the first credibility test. Sakana published a technical report at arxiv.org/abs/2606.21228 that provides a methodology baseline for third-party replication.

The above-272K-token pricing tier remains unspecified. Whether that rate makes Fugu competitive for long-context agentic workloads, which regularly exceed that threshold once scaffolding context, prior tool outputs, and intermediate results accumulate, will shape real-world adoption among the teams Fugu is explicitly targeting.

Sources