Sakana’s Fugu routes around Claude Fable 5 export limits with one OpenAI-compatible API

Last night, Sakana rolled out Fugu, a multi-agent orchestration system meant to deliver “frontier-level” AI performance through one OpenAI-compatible API, without forcing enterprises to bet everything on a single model provider. The pitch lands hard right where the market is nervous: Ha frames Fugu as a way to keep enterprise workflows running even when the most capable models get pulled.

That positioning traces directly to Anthropic’s move on June 12 to revoke public access to its most powerful models, Claude Mythos 5 and Claude Fable 5, following a US government export control order. In an X post today, Sakana’s CEO and co-founder David Ha said Fugu “dynamically orchestrates the world’s best models to tackle complex tasks,” and argues that “relying on a single company’s model for national infrastructure is a massive risk.” The crux is straightforward: instead of deploying one monolithic foundation model, Fugu dynamically routes tasks across a swappable pool of specialized AI agents, so restrictions on one model do not automatically shut down the entire system.

Under the hood, Sakana says Fugu is an LLM trained to call other LLMs in an agent pool, and it can even recursively call instances of itself. The technical release describes a lifecycle approach: when a complex request comes in, Fugu breaks the problem down, delegates sub-tasks to expert foundation models, verifies their work, and synthesizes the final answer. Sakana links this approach to learned coordination strategies grounded in two 2026 research papers, TRINITY and the Conductor, emphasizing autonomy in model selection and verification rather than hand-designed workflows.

To keep this usable for developers and enterprises, Sakana abstracts the multi-agent swarm behind a standard API endpoint. Users do not see the routing. Sakana explicitly states that which models Fugu selects, and how it coordinates them, are proprietary, and the documentation intentionally stays general, referring only to a “diverse pool of powerful models,” “multiple LLMs,” or “specialized models” without listing a specific count.

Where Fugu gets interesting for operational leaders is the control layer. Sakana offers an escape hatch for compliance-minded teams: developers can explicitly opt specific models or providers out of the Fugu routing pool, and users can opt out of having their prompts used for future training data. On availability and geography, Fugu is restricted from operating within the EU and EEA for now, while Sakana says it is working to align its black-box data routing architecture with GDPR regulations.

Fugu is also sold as two tiers with different performance and pricing behavior. Fugu is the high-speed, low-latency version designed for everyday tasks, positioned as a default engine for interactive chatbots and integrated directly into coding environments like Codex. Fugu Ultra is the flagship tier engineered for complex, high-stakes work such as AI research, cybersecurity analysis, and multi-step patent investigations. Sakana claims Fugu Ultra coordinates a deeper pool of experts and matches industry-leading monolithic models on scientific and reasoning benchmarks.

Sakana’s benchmark claims are part of the argument that this is not “only routing,” it is routing plus quality. For LiveCodeBench, an open-source benchmark of coding performance on regularly refreshed software problem-solving tasks, Sakana shows Fugu Ultra at 93.2 and Fugu at 92.9, compared with Claude Fable 5 at 89.8. On GPQA-D (Diamond), a test of 198 graduate-level multiple-choice questions across biology, physics, and chemistry, Sakana shows Fugu Ultra at 95.5 and another 95.5 score paired with Claude Mythos Preview at 94.6. The operational takeaway for executives is the redundancy angle: if multiple providers back the system, the stack can route around an outage or sudden regulatory restriction, aiming to maintain uptime.

Pricing and deployment are designed to be predictable at scale, at least for the Ultra tier. Fugu is available immediately in most regions, with a temporary exception for the EU and EEA, through subscription tiers and pay-as-you-go plans. Sakana lists Standard at $20/month for lightweight workflows, Pro at $100/month providing 10x standard usage, and Max at $200/month offering 20x usage for continuous, long-running tasks. For the subscription rollout, Sakana is also offering a free second month for users who subscribe to any tier by July 31, 2026. VentureBeat notes it could not find the exact token coverage for those subscription allowances and reached out to Ha for more information.

For consumption-based enterprise scaling, the pay-as-you-go setup is where procurement teams will focus. For high-priority environments, requests served under the consumption-based model are served at a higher priority than those from monthly subscription plans. Sakana says standard Fugu charges a single rate of the highest-tier underlying model involved in a query, without stacking multi-agent fees. For Fugu Ultra (fugu-ultra-20260615), pricing is fixed per one million tokens: $5 input, $30 output, and $0.50 cached input. Rates increase for extreme workloads utilizing context windows above 272K tokens to $10 input, $45 output, and $1.00 cached input. VentureBeat also includes a Frontier AI Model API Pricing Snapshot, which suggests Fugu Ultra could be expensive relative to some single-model APIs depending on context length and provider rates.

Here is the second-order risk executives should absorb: orchestration models are effectively a new concentration point. If the customer-facing API becomes the control plane for multi-model routing, then procurement, security, and legal teams will need to treat that orchestration vendor as mission critical. Sakana’s pitch is that Fugu reduces concentration risk from any one underlying provider, especially under export-control volatility. The counterweight is that you now depend on Sakana’s proprietary coordination patterns and availability. For founders building enterprise AI, investors evaluating platform defensibility, and CIOs scaling production deployments, the question becomes less “which single model is best,” and more “which orchestration approach keeps systems reliable when model access, policies, or providers change overnight.”