Sakana AI releases Fugu, an LLM that orchestrates a pool of expert models and matches frontier-level performance through a single API

Sakana AI launched Fugu on June 22, 2026, a new kind of AI model designed to orchestrate a dynamic pool of large language models through a single OpenAI-compatible endpoint. Unlike a traditional single model, Fugu is itself an LLM trained to decide—on a per-request basis—whether to answer directly or route the task to one or more specialized expert models, then synthesize their outputs into a unified response.

What's new

Fugu ships in two variants:

• Fugu: Optimized for balanced performance and low latency. Suited for everyday coding, code review, and interactive chat workflows. Designed to act as a drop-in replacement for current API-based model calls.
• Fugu Ultra (fugu-ultra-20260615): Tuned for maximum accuracy on demanding, multi-step tasks—Kaggle competitions, cybersecurity analysis, multi-hour research runs. Priced at $5/M input tokens, $30/M output tokens, and $0.50/M cached input tokens.

Both variants are generally available today via a single OpenAI-compatible API. Subscription tiers cover everyday use; pay-as-you-go plans handle enterprise-scale workloads. As a launch promotion, Sakana is offering a free second month for users who subscribe by July 31, 2026.

The underlying architecture draws on Sakana's TRINITY and Conductor research papers. A 7-billion-parameter conductor model learns which specialist models to activate, how to coordinate communication between agents, and how to combine their outputs. The agent pool—which includes models from multiple providers—is designed to be swappable, so a single provider disruption routes around the gap rather than failing the request.

Context

Sakana AI was founded by former Google DeepMind and Google Brain researchers with roots in the original Transformer paper. The lab raised a $135M Series B at a $2.65B valuation in late 2025. Prior launches included an Ultra Deep Research agent for 100-page reports and work on Continuous Thought Machines.

Fugu was in beta from April 25, 2026 before today's general availability launch.

Benchmark results peg Fugu Ultra at or near Anthropic's Fable 5 and Mythos Preview across coding, reasoning, scientific, and agentic evaluations.

Why it matters

Fugu represents a structural bet that multi-model orchestration can match single-model frontier performance, without training a trillion-parameter model from scratch. The approach lowers the cost of reaching frontier capability to building a strong coordinator and assembling a pool of best-of-breed models.

The export control angle is also relevant. U.S. export controls have restricted access to American frontier models in certain geographies since June 12, 2026. A Japanese lab building a frontier-equivalent system via orchestration creates a viable alternative for teams where American models are restricted.