While most of the AI industry races to build one ever larger model, a Tokyo startup has launched a system that takes the opposite approach, an AI whose main skill is directing other AIs. Sakana AI released Fugu on June 22, 2026, a model that behaves less like a competitor to the frontier systems and more like a conductor standing in front of them. Send Fugu a single request through one API, and it decides on the spot whether to answer the question itself or to assemble a team of specialized models, hand each one a piece of the work, check what they return, and merge the results into a single answer.
The company behind it carries unusual weight in the field. Sakana AI is based in Tokyo and was co-founded by researchers including Llion Jones, a co-author of “Attention Is All You Need,” the 2017 paper that introduced the Transformer architecture that nearly every modern AI model is built on, alongside David Ha, formerly of Google Brain and Stability AI. That lineage matters because Fugu is a bet against the very scaling race that the Transformer helped set off, an argument that the next gains may come not from training a bigger brain but from coordinating the brains that already exist.
Fugu ships in two versions aimed at different jobs. The standard Fugu is tuned for speed and low latency, built for everyday coding, chat, and quick review work where a fast answer matters more than exhaustive depth. Fugu Ultra is the heavyweight, designed for hard, multi-step problems such as AI research, reproducing scientific papers, cybersecurity assessments, and patent search, where the system can spend longer assembling and verifying the work of several models. Both are reached through the same unified, OpenAI-compatible endpoint, with subscription plans for regular users and pay-as-you-go pricing for high-volume and enterprise workloads.
The performance claims are bold. Sakana says Fugu Ultra matches Anthropic’s Fable 5 and Mythos Preview, two of the strongest current models, on its published tables for the hardest engineering, scientific, and reasoning benchmarks. The company goes further, claiming Fugu beats Gemini 3.1 Pro, Opus 4.8, and GPT-5.5 on tasks including automated research, mechanical design, and financial forecasting, and on coding tests such as SWE-Bench Pro and TerminalBench. What makes those numbers notable is the method behind them. Sakana reached them without training a new frontier model of its own, instead training a roughly seven billion parameter orchestrator, a small “conductor” model that learns which outside systems to call, how they should talk to one another, and how to combine their output, drawing on two of the company’s papers presented at the ICLR 2026 conference.
The claims deserve a measure of caution. Independent verification is still pending, and orchestration systems have historically underperformed their benchmark scores in real use, a familiar gap for systems introduced with strong published numbers. The benchmark tables are Sakana’s own, and orchestration adds moving parts, the latency of calling several models, the cost of running them, and the risk that the conductor makes a poor choice, that a single trained model does not carry. The idea is elegant, but whether it holds up across the messy variety of real tasks is exactly what the coming weeks of outside testing will decide.
The architecture also answers a problem that has grown sharper as governments tighten access to advanced AI. Because Fugu’s pool of expert models is swappable rather than fixed, the system is designed to route around a provider that becomes unavailable, so if export controls or other restrictions cut off one model, the conductor can lean on others. Sakana frames this as a hedge against both vendor lock-in and the political fragmentation of the AI supply chain, a way to keep a service running even as the map of which models can be used where keeps shifting.
Fugu’s deeper bet is conceptual. Most of the industry treats the model as the product and scale as the path forward, pouring resources into training single systems that are larger and more capable than the last. Fugu treats orchestration as the product, wagering that a smart layer sitting above a field of strong models can match or beat any one of them by always using the right tool for each part of a job, even calling itself recursively when a problem needs to be broken down further. If that bet pays off, the most valuable position in AI may shift from owning the biggest model to conducting the best ensemble, and a Tokyo lab founded by one of the Transformer’s own architects will have made the first serious case for it.
Related: Google’s Gemini 3.5 Pro preview · OpenAI’s GPT-5.5-Cyber tops the CyberGym benchmark
The big picture: The AI Model Race (2026): who’s winning the frontier contest