When the US government ordered Anthropic to halt access to its top AI models for foreign nationals on June 12, it created something unusual in the AI market: a capability gap that frontier competitors are now rushing to fill.

Tokyo-based Sakana AI launched Fugu Ultra on June 22, 2026, positioning it directly into that gap. The company describes Fugu as “delivering frontier capability without the risk of export controls” — and the benchmarks back up the claim.

On SWE-Bench Pro, the industry’s most demanding software engineering evaluation, Fugu Ultra scored 73.7%. That puts it ahead of Claude Opus 4.8 (69.2%), GPT-5.5 (58.6%), and Gemini 3.1 Pro (54.2%). It also leads on Humanity’s Last Exam and four other coding benchmarks.

How Fugu Actually Works

Fugu is not a single model in the traditional sense. It is a multi-agent orchestration system delivered as a single API endpoint. The system is itself a language model trained to call a swappable pool of other frontier LLMs — including instances of itself — and route each sub-task to whichever model is best suited to handle it.

The result is that a single API call to Fugu Ultra might internally coordinate several specialist models: one for reasoning, one for code generation, one for verification. The user sees one response. The complexity is invisible.

This architecture has two immediate implications for businesses:

First, you are not locked into any single model provider. If one model improves or becomes unavailable, Fugu updates its pool without requiring changes to your integration.

Second, the benchmark performance reflects the ceiling of what the underlying model pool can achieve — not the average. Tasks go to the strongest available model for that task type.

Fugu comes in two variants. The standard Fugu handles latency-sensitive everyday work. Fugu Ultra is tuned for complex, multi-step problems where accuracy matters more than speed. Pricing for Fugu Ultra is $5 per million input tokens and $30 per million output tokens, rising to $10/$45 once context exceeds 272K tokens.

The API is fully OpenAI-compatible. Teams already using GPT, Gemini, or Claude via API can swap in Fugu Ultra with minimal code changes.

The Export Control Context

This launch comes at a specific moment. On June 12, the US Department of Commerce ordered Anthropic to suspend access to Fable 5 and Mythos 5 for all foreign nationals — including foreign nationals working at companies inside the United States. Anthropic’s top-tier models went dark for a significant portion of the global market overnight.

Sakana AI’s spokesperson told TechCrunch that the timing of Fugu’s release was “entirely coincidental.” The company has not hesitated to capitalise on the moment regardless. Its marketing explicitly calls out the export control risk as a reason to consider Fugu.

Chinese firm 360 has also moved quickly, unveiling Tulongfeng, an AI model it says can compete directly with Anthropic’s Mythos — a move that sits at the intersection of commercial competition and geopolitical positioning.

The Anthropic export ban is not expected to be permanent. The US government allowed a limited rollout of GPT-5.6 Sol to approximately 20 companies on June 26. But the episode has exposed something that many enterprise AI buyers had not previously considered: frontier AI access can be cut off by government order.

What This Means for Business

For most businesses using AI for productivity, content, and internal automation, the Anthropic ban was an inconvenience rather than a crisis. But for companies that had built workflows or products around Claude Fable or Mythos specifically, it was a hard lesson in vendor concentration risk.

The Fugu launch matters for three reasons beyond its benchmark numbers.

The first is optionality. If your AI stack depends on one provider staying accessible, one government directive can break it. An orchestration layer that swaps underlying models insulates you from that risk. Fugu is one example of that architecture; building your own multi-provider routing is another.

The second is pricing leverage. When multiple models compete at the frontier, buyers benefit. Fugu Ultra at $5/$30 per million tokens undercuts Anthropic’s and OpenAI’s highest tiers. The market is getting more competitive at the top end, and that compression is good for enterprise buyers.

The third is that the definition of “frontier” is no longer American by default. Sakana AI’s research background is rigorous — Fugu builds on ICLR 2026 papers, including Trinity and Conductor, both on learned model orchestration. The frontier AI field is genuinely global now, and that changes what vendor evaluation looks like.

For enterprise data teams and AI builders at Enterprise DNA, the practical takeaway is to architect for substitutability from the start. Whether you use Claude, GPT, Gemini, or Fugu, your application logic should sit above the model layer — not be woven into it. The companies that did this in early 2026 felt almost nothing when the export ban hit. The ones that hadn’t were scrambling.

The export ban episode will not be the last of its kind. Building model-agnostic infrastructure is no longer theoretical best practice. It is basic operational resilience.