On June 22, 2026, Tokyo-based Sakana AI launched public access to Fugu and its advanced version, Fugu Ultra. Rather than being just another large language model, Fugu acts as an orchestrator—a specialized model that receives requests via a single OpenAI-compatible endpoint and determines whether to handle them solo or assemble a team of third-party frontier models like GPT-5.5, Claude, and Gemini, or even recursively call itself. Role allocation, verification of intermediate steps, and the synthesis of the final response are all managed internally, requiring no additional logic in the user's code.
This concept did not emerge in a vacuum. The first half of 2026 highlighted the vulnerability of relying on a single provider, particularly when Anthropic was forced by U.S. export mandates on June 12 to shut down public access to its most powerful models, Claude Fable 5 and Mythos. For corporations and governments that have integrated third-party APIs into critical infrastructure, such overnight disruptions are no longer just a hypothetical risk. Sakana’s response is pragmatic: rather than chasing parameter counts, the focus is on assembling the best possible team from existing resources and bypassing missing links. Since the model pool is interchangeable, the system simply routes tasks around unavailable models like Fable or Mythos.
The foundation of this technology lies in two papers accepted for ICLR 2026. TRINITY details an advanced coordinator that assigns roles like Thinker, Worker, and Verifier to orchestrate multi-step interactions (arXiv:2512.04695). Conductor is a model trained via reinforcement learning to generate coordination plans in natural language (arXiv:2512.04388). Instead of following hard-coded workflows, Fugu learns collaborative patterns, granting it the flexibility to integrate new models as soon as they are released.
There are two available versions. The standard Fugu model balances quality and latency, serving as a workhorse for coding, reviews, and chatbots—including those within Codex—while allowing users to exclude specific models or providers to meet privacy and compliance standards. Fugu Ultra is tailored for long, multi-step tasks where monolithic models often lose context or accumulate errors, such as reproducing scientific research, deep code analysis, cybersecurity, and patent or literary searches.
Regarding the numbers, Sakana compared Fugu against the very models it orchestrates, and the orchestrator outperformed each individual participant on most metrics. Fugu Ultra currently leads on major coding benchmarks and several scientific tests; the lab even claims it matches the performance of Fable 5 and Mythos Preview, despite neither being in the current pool due to their lack of public availability. Since all third-party results are pulled from the providers' own reports, they should be viewed as claimed figures rather than independently verified data.
The strategic advantages are threefold. First, it reduces vendor lock-in by allowing models from different companies to be mixed based on cost, speed, and compliance. Second, it provides built-in redundancy, enabling the system to bypass outages or new restrictions from any single provider. Finally, the billing structure is more transparent than traditional multi-agent systems; when multiple agents are active, the fee is calculated at the rate of the most expensive model in the chain rather than compounding.
Access is provided via a unified API. Subscriptions are offered in three tiers: Standard for $20 per month, Pro for $100 (with ten times the limit), and Max for $200 (twenty times the limit), with a promotion running through the end of July that offers the second month for free. For heavy workloads, a pay-as-you-go option is available; Fugu Ultra costs $5 per million input tokens and $30 per million output tokens, with higher rates for context windows exceeding 272K. The technical report is available on the lab's GitHub, while the product and console can be found at sakana.ai/fugu and console.sakana.ai. As a significant caveat, the service is currently unavailable in the EU and EEA at launch, as Sakana cites ongoing work toward GDPR compliance.
