Sakana AI released Fugu Max, a model that coordinates multiple smaller AI systems to answer questions rather than relying on a single large model, achieving competitive performance at significantly lower cost. The Tokyo-based research lab, founded by three former Google researchers including Llion Jones (a co-author of the original transformer paper), is betting that the future of AI lies in orchestrated networks of specialized systems rather than monolithic models.

How Does Fugu Max Actually Work?

Fugu Max operates through an intelligent routing system that decides which model should handle each task. When you send a request to Fugu Max, you are not talking to a single neural network. Instead, you are communicating with a coordinator that analyzes your query and assigns it to the cheapest model capable of handling it well.

The coordination system itself is grounded in two research papers from ICLR 2026. One framework, called TRINITY, assigns roles like Thinker, Worker, and Verifier across multiple language models over several turns. Another system, called the Conductor, was trained using reinforcement learning to discover natural-language coordination strategies, meaning the system learned how to delegate work rather than following a hand-written set of rules. The system can even call instances of itself recursively to break down complex problems.

For developers, the appeal is straightforward: all of this complexity hides behind a single application programming interface (API) endpoint. You get the benefits of a multi-agent system with expert models routed by task, without having to build and maintain that system yourself.

What Are the Performance Claims and Benchmarks?

Sakana reports that Fugu Max achieves the best overall score on six benchmarks, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. The company claims it "expands the cost-performance Pareto frontier on seven out of ten benchmarks," with performance close to elite models at two to six times lower cost.

Sakana

Fugu Max's sibling model, Fugu Ultra v2, launched the same day with more dramatic numbers. On Chartography, a visual-reasoning benchmark, Ultra v2 scored 48.3 compared to Opus 5 at 27.3 and Fable 5 at 29.5. On DeepSWE, a real-world software-engineering benchmark, it achieved 74.3, which Sakana says beats models costing three to five times more per token.

However, three important caveats apply. First, SWEFish is Sakana's own internal benchmark reflecting its own coding challenges, not a third-party test. Second, the strongest proprietary models are not in the pool; Sakana explicitly states that Fable 5, Fable 5.1, and GPT-6-Astra are not among the models Fugu orchestrates. Third, routing is opaque by design; Sakana does not tell you which underlying models answered a given query, so you cannot audit the mix per request.

What Is the Pricing Structure?

Fugu Max uses a flat-rate pricing model that holds regardless of context length. Input costs $10 per million tokens, and output costs $30 per million tokens. Web search and fetch operations cost $0.007 per call. The flat-rate design means a 900,000-token prompt costs the same per token as a short one, which is unusual and useful if you process large documents.

Sakana also offers three subscription tiers that include access to all Fugu models:

• Standard Plan: $20 per month with baseline allowance
• Baseline Plan: $100 per month, providing 10 times the Standard allowance
• Premium Plan: $200 per month, providing 20 times the Standard allowance

Two factors affect the real cost. First, because orchestration tokens are metered, a request that routes to several agents burns more tokens than a single-model call would, so the low per-token rate does not always mean a low per-answer bill. Sakana softens this by promising it "never stacks fees" and charges "a single rate based on the top-tier model involved." Second, caching significantly reduces costs in practice; on OpenRouter's live panel in Fugu Max's first days, the weighted-average effective input price came in around $0.61 per million against the $2 list price, driven by a 79.4% cache-hit rate.

What Are the Key Technical Specifications?

Fugu Max supports a 1-million-token context window, which means it can process roughly 1 million words at once, with text, image, and PDF inputs and text output. Users can configure reasoning effort at high, xhigh, and max levels. The model supports function calling, structured outputs, and built-in web search and fetch capabilities.

One important limitation: the agent pool is fixed. Unlike the base Fugu model, you cannot opt specific models in or out of Fugu Max from the console. This design choice prioritizes simplicity over customization.

How Has the AI Community Reacted?

The Fugu Max launch was quiet online, but the broader debate around Fugu's orchestration approach has been loud since the model's general availability launch in June. The dominant critique centers on the value proposition: "Why pay for a router?" One commenter on Hacker News captured the skepticism, noting that paying subscription fees to multiple AI providers and then paying Sakana to coordinate them all seemed redundant.

However, there was a fair counterpoint from other technical observers. One developer noted: "Kind of shocking, a model comes out that beats mythos and offers a reasonable price and it gets downvoted? It's a genuine accomplishment to hit those benchmarks and offer a reasonable plan".

The real question for technically minded users remains whether building and maintaining a custom orchestration system with cheap open-weight models themselves would deliver better value than using Fugu Max. For many developers, the answer depends on whether they have the engineering resources to build and maintain such a system versus paying for a managed solution.