Sakana AI Fugu Max and Fugu Ultra v2 Bet on Routing Over Frontier Scale

Sakana AI has shipped Fugu Max and Fugu Ultra v2, a pair of orchestration systems that route each request across a pool of open and specialized models instead of running one large model for every task. The Sakana AI Fugu Max and Fugu Ultra v2 release, dated September 11, splits a single learned routing architecture into a cost-first product priced at $2.00 per million input tokens and $6.00 per million output tokens, and a capability-first version that Sakana says leads on visual reasoning and software engineering results without a proprietary frontier model in its pool. Both are available immediately through the Sakana AI API.

Neither system is a foundation model. Fugu is a learned orchestrator that picks which model in its pool should handle a request, then returns the answer behind one OpenAI-compatible endpoint. The two releases share that core architecture and differ in how it is tuned and what it is permitted to route to. Sakana frames Fugu Max as expanding the Pareto efficiency frontier outward and Fugu Ultra v2 as pushing it upward.

Sakana AI Fugu Max and Fugu Ultra v2: Two Products on One Router

Fugu Max is the cost play. Sakana prices it 40–60% below Sonnet 5 and GPT-5.6 Terra, and it draws on the largest pool of open and specialized models the company has assembled, including NVIDIA's Nemotron family under a collaboration with NVIDIA. Sakana reports that Fugu Max leads Terminal-Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench and SWEFish.

Fugu Ultra v2 targets complex reasoning, visual interpretation and software engineering, with a training cutoff of 20260828. Sakana reports 48.3 on Chartography against 27.3 for Opus 5, and 74.3 on DeepSWE. Its pool deliberately excludes Fable 5, Fable 5.1 and GPT-6-Astra.

The split is the product decision. Earlier Fugu messaging sold one system that matched frontier output through autonomous model selection, and the new line turns that into two tiers, one tuned for the token bill and one for the hardest tasks. The version numbering also points to iteration rather than a new family, since Ultra v2 supersedes a first-generation Ultra.

The Cost Case and the Lock-In Case

Pricing is where the announcement bites. At $2.00 per million input tokens, Fugu Max undercuts the flagship tier of the major labs by a wide margin, and Sakana's argument is that the saving does not require accepting weaker output on the tasks it targets. For teams running high-volume agent workloads such as document extraction, batch research, code review and test generation, input tokens dominate the bill, and a 40–60% reduction compounds through every retry and every tool call in a chain.

The arithmetic is easy to run for a large buyer. A workload moving 100 million input tokens a month costs roughly $200 at Fugu Max rates, while the implied competitor range of about $3.33 to $5.00 per million input tokens that the stated discount produces would put the same traffic at $333 to $500. Output pricing at $6.00 per million follows the same logic at a smaller volume.

Pool design carries a second argument that matters more to procurement than to engineering. Because the router can swap models in and out, Sakana positions the architecture as insulation from API revocations and abrupt service cutoffs, and as protection from a single supplier holding pricing power over a production stack. A long-term contract signed against one proprietary endpoint inherits that vendor's future terms. A routing layer turns the same dependency into a configuration change, and the OpenAI-compatible interface keeps migration close to a base-URL swap.

The trade-off is that the quality ceiling is set by whatever sits in the pool at a given moment. Leaving Fable 5, Fable 5.1 and GPT-6-Astra out removes lock-in and also removes the strongest proprietary systems from reach, so Ultra v2's scores rest on how well the router combines weaker components.

Where the Trade-Offs Land

Latency is the clearest cost. On OpenRouter, Fugu Max's median response time is about 5.4 seconds, a direct result of the routing step and the multi-model work behind it. In a multi-step agent loop that delay multiplies, and a twenty-call chain carries close to two minutes of overhead before any model produces a token. That rules the system out of interactive chat, autocomplete and anything where a user waits on the first token. Asynchronous pipelines absorb the delay without noticing.

Benchmark claims need the usual discount. A 48.3 to 27.3 gap on Chartography is one visual reasoning measure, and it does not establish general superiority over Opus 5 in reasoning or coding. The same caution applies to the six leaderboard wins credited to Fugu Max: a strong signal across task types, not a blanket result. Both sets of scores describe a pool that changes as open-weight models are updated.

Cost predictability is the quieter problem. Because the router decides which model handles which request, a customer's bill depends on how their own traffic is classified, and a shift in task mix can move unit economics without any change in list price. Buyers accustomed to a flat per-token rate should model their own workloads before trusting a headline price. For teams already running open-weight models in-house, the real comparison is not Sakana's list price but the fully loaded cost of their own inference stack, including the engineering time spent on routing logic they would otherwise build.

Sakana has published no per-token price for Fugu Ultra v2, which leaves the central question of what peak capability costs relative to the cheap tier unresolved.

What the Orchestration Layer Changes for the Market

Fugu Max is also carried by the routing marketplace OpenRouter, which places it alongside that aggregator's other endpoints rather than in a category of its own. What separates Sakana's approach is that the routing decision is learned rather than fixed by price or context-length rules, and that the company publishes head-to-head wins against named proprietary models.

The business model behind both releases is arbitrage rather than training. Sakana earns its margin by choosing cheaper models per task and charging for the routing decision, so the spread widens if open-weight inference keeps getting cheaper and narrows if proprietary labs cut mid-tier prices to defend volume. That makes Sakana's economics sensitive to decisions taken by the same vendors it routes around.

The NVIDIA collaboration cuts both ways. Adding the Nemotron family to the pool gives NVIDIA a route into enterprise stacks that standardize on OpenAI-compatible interfaces, and it gives Sakana a large-vendor partner for its cost tier. It also ties part of Fugu Max's quality to another company's release schedule, the same dependency the router is meant to soften for its own customers.

Benchmark coverage says something about where the system is meant to work. Terminal-Bench 2.1, GDP.pdf, AA-LCR, AutomationBench and SWEFish are agentic, terminal-driven, document-heavy and software tasks, the workloads that generate the largest token bills. Sakana is not claiming a general-purpose lead. It is claiming the high-volume middle of enterprise AI.

Why this matters

Sakana is selling orchestration as a product category, and if routing can hold quality near the frontier while halving the token bill, the pressure lands on mid-tier proprietary models that charge near-flagship prices for work an open pool can absorb. For buyers, the sensible move is a narrow pilot: run one high-volume, latency-tolerant workload through Fugu Max and compare the bill and the error rate against the model it would replace. The numbers to watch are Ultra's price and whether the pool composition, and with it the scores, survives the next round of open-model releases.

Related Articles

• AI Model Fatigue Grips Buyers as Four Frontier Labs Ship in One Week
• Moonshot AI's Kimi K3 Open Weights Reshape the Frontier Model Economics
• AI Business Roundup: Capital Floods Compute as Regulators Tighten

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.