Sakana AI Launches Fugu: A Single API for Multi-Agent Orchestration

AIDeveloper44 Team

The End of Monolithic Dependency

For the past few years, the AI industry's progress has been driven largely by brute-force scale: building giant, monolithic models trained on ever-larger amounts of data. However, hard, real-world tasks require a multitude of specialized knowledge and skills that surpass the capabilities of any single benchmark-topping monolith. To solve this, Sakana AI has officially launched Sakana Fugu, an innovative system that delivers full multi-agent orchestration as a single foundation model.

Announced in late June 2026, Sakana Fugu fundamentally shifts the AI paradigm. Instead of relying on a single, isolated monolith, Fugu dynamically orchestrates the world's best expert models to tackle complex, multi-step workflows. By packaging this collective intelligence into a single API, developers can achieve frontier-level performance without being shackled by single-vendor dependency.

AI Sovereignty and Geopolitical Resilience

Sakana AI explicitly designed Fugu to address the growing operational and geopolitical risks of relying on a single provider. Recent industry disruptions—such as export controls imposed on Anthropic's Fable 5 and Mythos models—have highlighted how critical infrastructure can become vulnerable overnight due to shifting regulatory boundaries. Fugu mitigates this by acting as a dynamic router. If a single provider goes down or restricts access, Fugu automatically routes around the disruption using an underlying pool of completely swappable agents. This approach provides a practical, resilient blueprint for AI sovereignty.

Meet the Models: Fugu and Fugu Ultra

At launch, the Sakana Fugu ecosystem is divided into two distinct tiers tailored to different workloads, both accessible through a unified endpoint:

• Fugu: Designed as the ideal default for everyday tasks. Fugu balances strong performance with low latency. It is highly optimized for standard coding tasks, code reviews (like dropping into tools such as Codex), and powering highly responsive interactive chatbots.
• Fugu Ultra: Built for the hardest, most demanding workflows. Fugu Ultra coordinates a deeper pool of expert agents (routing between one to three models depending on the problem) to maximize quality on high-stakes tasks. Early adopters are already leveraging Fugu Ultra for Kaggle competitions, automated ML research, cybersecurity assessments, and patent investigations.

Frontier-Level Benchmarks

When evaluated against the leading 2026 frontier models—including Gemini 3.1 Pro, Opus 4.8, and GPT-5.5—Fugu Ultra proves that coordinated collective intelligence can outmatch single massive models. According to Sakana AI's quantitative results, Fugu Ultra achieved staggering numbers across rigorous scientific and coding benchmarks:

• SWE Bench Pro: Fugu Ultra scored 73.7, crushing GPT-5.5 (58.6) and Opus 4.8 (69.2).
• LiveCodeBench: Fugu Ultra achieved 93.2, narrowly beating the standard Fugu (92.9) and easily surpassing Opus 4.8 (87.8).
• GPQA Diamond: Both Fugu and Fugu Ultra maxed out at 95.5, edging past GPT-5.5 (93.6).

Beyond raw numbers, Sakana AI showcased impressive qualitative victories. In a test using AutoResearch (an autonomous ML framework), Fugu Ultra successfully ran 123 experiments over 14 hours to improve a small GPT's training recipe, finding optimizations across batch size, model depth, and learning rates to achieve the best bits-per-byte (BPB) validation score against all baselines.

The Science: TRINITY and The Conductor

Fugu's seemingly magical ability to orchestrate isn't driven by rigid, hand-coded scripts. The system is built upon two landmark ICLR 2026 papers published by Sakana AI:

TRINITY: An Evolved LLM Coordinator - This research introduces a lightweight, evolved coordinator that governs multiple LLMs across several turns. It dynamically assigns Thinker, Worker, or Verifier roles, adaptively delegating work based on the nature of the math, coding, or reasoning task.

Learning to Orchestrate Agents in Natural Language with the Conductor - Trained via reinforcement learning, the Conductor discovers natural-language coordination strategies. It creates sophisticated communication patterns and focused prompts that allow a diverse LLM pool to systematically outperform individual models on difficult benchmarks.

Developer Experience: An OpenAI-Compatible API

Despite its complex multi-agent architecture under the hood, integrating Fugu is remarkably simple. The platform exposes a fully OpenAI-compatible API. Developers can swap out their base URL to https://api.sakana.ai/v1 (or their console-provided endpoint), use standard OpenAI SDKs, and begin querying.

For optimal results—especially regarding tool use, multi-step reasoning, and web searches—Sakana AI strongly recommends using the newer /v1/responses API rather than standard Chat Completions. Fugu includes built-in OpenAI-compatible tools, such as web_search, natively within the array.

Transparent Orchestration Billing

Because Fugu Ultra actively calls multiple agents, its token usage differs from standard models. Sakana AI has introduced transparent token tracking in their API responses. Developers will see standard input_tokens and output_tokens, but the JSON payload also includes input_tokens_details.orchestration_input_tokens and output_tokens_details.orchestration_output_tokens. These represent real token usage generated by the system's internal routing and deliberation. While these tokens count toward the final cost of the request, they are priced identically to standard input and output tokens, keeping billing straightforward.

Looking Ahead

Sakana Fugu represents a massive leap forward in how we deploy and scale AI. By turning multi-agent orchestration into a learned, accessible foundation model, Sakana AI is charting a path away from monolithic dominance toward a collaborative, resilient ecosystem. Fugu is generally available now, with both subscription tiers for standard use and pay-as-you-go pricing for enterprise scale.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe