# Sakana AI's Fugu: A Router Built to Survive a Model Being Switched Off

Ten days into a government-ordered shutdown of two Anthropic models, Sakana AI shipped Fugu, a router that treats every underlying model as swappable behind one endpoint.

Author: J.A. Watte
Published: July 6, 2026
Source: https://jwatte.com/blog/sakana-fugu-model-router/

---

In the middle of June 2026, one of the most capable AI systems in the world stopped answering. It did not break, and its maker did not change its mind. A government told the company to switch it off. For roughly nineteen days, two of Anthropic's frontier models went dark. Ten days into that blackout, a startup in Tokyo shipped a product built around exactly that failure, a router meant to keep working when the model underneath it gets pulled. They named it after a pufferfish.

## The nineteen days a frontier model went dark

On June 12, 2026, Anthropic said it received a directive from the US government at 5:21pm ET ordering it to suspend all access to Claude Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including its own foreign national employees ([Anthropic](https://www.anthropic.com/news/fable-mythos-access)). The same statement said access to all other Anthropic models would not be affected ([Anthropic](https://www.anthropic.com/news/fable-mythos-access)). So this was not a company retiring an old model. It was an outside party reaching into a live product and turning two of its best engines off.

The stated basis was national security. Anthropic connected the action to what it called a potential narrow, non-universal jailbreak, which it described as asking the model to read a specific codebase and fix any software flaws, and it disputed the idea that this warranted recalling a commercial model deployed to hundreds of millions of people ([TechCrunch](https://techcrunch.com/2026/06/12/anthropics-safety-warnings-may-have-just-backfired-the-government-has-pulled-the-plug-on-its-most-powerful-ai/)). The language is Anthropic's own, and it reads as a company that felt the response was out of proportion to the problem it described.

On paper, the order aimed at foreign nationals. In practice, TechCrunch's Connie Loizos reported, it forced Anthropic to disable both models for all users worldwide, not only the foreign nationals the export-control order was nominally aimed at ([TechCrunch](https://techcrunch.com/2026/06/12/anthropics-safety-warnings-may-have-just-backfired-the-government-has-pulled-the-plug-on-its-most-powerful-ai/)). Al Jazeera described the same move: the administration told the company to bar all foreign nationals, including its own staff, from the models, citing unspecified national security concerns ([Al Jazeera](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says)).

Whether US persons were legally required to lose access, or whether Anthropic disabled the models globally to stay on the safe side, is not something the public statement settles. What is clear is the result. Two frontier models, gone for everyone, with almost no notice. That gap between the narrow legal target and the broad practical effect is the part builders should sit with, because it is the part a contract cannot easily protect you from.

On June 30, 2026, the Commerce Department lifted the controls, ending what multiple outlets called a nineteen-day shutdown ([CNBC](https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html)). Anthropic confirmed it on its official account, saying the Department of Commerce had lifted export controls on Claude Fable 5 and Mythos 5 and that it would begin restoring access the next day ([Anthropic](https://x.com/AnthropicAI/status/2072106151890809341)). Access came back around July 1 ([Al Jazeera](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says)).

Anthropic did not get the two models back for free. According to Al Jazeera, Commerce Secretary Howard Lutnick's letter said the company would no longer require an export licence because it had agreed to three commitments: to proactively detect and address security risks tied to the models, to work with the government on standards for upcoming models, and to inform the government of any "malicious activity" ([Al Jazeera](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says)). It drew named pushback too. Tanishq Abraham, formerly a research director at Stability AI, asked whether the precedent means "Does the US government need to approve every frontier model release?", tying it to OpenAI's GPT-5.6 launch, while Francesco Bailo of the University of Sydney said officials likely "overreacted" ([Al Jazeera](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says)).

## What Sakana shipped in the middle of it

On June 22, 2026, roughly ten days into the shutdown, Sakana AI launched Fugu ([Sakana](https://sakana.ai/fugu/), [MarkTechPost](https://www.marktechpost.com/2026/06/22/sakana-ai-launches-sakana-fugu-an-orchestration-model-that-routes-tasks-across-a-swappable-pool-of-frontier-llms/)). Its product page carries the line "A Multi-Agent System, Delivered as One Model" ([Sakana](https://sakana.ai/fugu/)), while the release page reaches for something grander: "One Model to Command Them All" ([Sakana](https://sakana.ai/fugu-release/)). The names promise the same thing, one interface with many models underneath it.

It helps to be precise about what Fugu is, because the marketing invites a misread. Fugu is not a single new frontier model that happens to be very good. It is an orchestrator. Sakana describes a system that dynamically coordinates a pool of powerful models and exposes all of them through one OpenAI-compatible API ([Sakana](https://sakana.ai/fugu/)). VentureBeat drew the distinction cleanly in its reporting: Sakana trained a coordinator rather than train a new large language model ([VentureBeat](https://venturebeat.com/orchestration/no-claude-fable-5-no-problem-sakana-achieves-frontier-performance-with-new-fugu-multi-model-auto-synthesis-system)). MarkTechPost called it an orchestration model that routes tasks across a swappable pool of frontier LLMs ([MarkTechPost](https://www.marktechpost.com/2026/06/22/sakana-ai-launches-sakana-fugu-an-orchestration-model-that-routes-tasks-across-a-swappable-pool-of-frontier-llms/)).

The coordinator's job is selection, delegation, verification, and synthesis. It decides which model or models should take a request, hands off the work, checks what comes back, and assembles a single answer. Sakana offers two tiers, Fugu and Fugu Ultra ([Sakana](https://sakana.ai/fugu/)). Secondary coverage has attached a specific parameter count to the trained coordinator, but Sakana's own pages do not state one, so any such number should be treated as unconfirmed.

There is an irony worth naming. The two Anthropic models at the center of the shutdown are not in Fugu's own pool. Sakana says so directly in a benchmark footnote, noting that neither of them is in Fugu's agent pool because they are not publicly accessible ([Sakana](https://sakana.ai/fugu/)). Fugu Ultra is nonetheless positioned as standing shoulder to shoulder with leading models like Anthropic's Fable 5 and Mythos Preview ([Sakana](https://sakana.ai/fugu-release/)), models it measures itself against without being able to call them.

## The risk it is really selling against

Step back from Fugu specifically and the interesting thing is the shape of the risk it names. Most continuity planning for software assumes the failure is technical: a server falls over, a region goes offline, a dependency ships a bad release. The Anthropic episode is a different category. The model did not fail. It was working fine. It was removed by a third party for reasons that had nothing to do with the customers who lost access, and the removal was total and fast. You cannot patch your way around that, and you cannot call support and ask for an exception, because the party that pulled the plug is not your vendor.

That is the risk a router addresses, at least in principle. If your application talks to an abstraction that can reach several models, then any one of them going dark becomes a routing decision rather than an outage. The value is not that Fugu is smarter than the models it calls. The value, if the framing holds, is that no single model is load-bearing. Whether a given router actually delivers that, without quietly degrading quality when it swaps engines, is exactly the thing to test before trusting it, because swapping one model for another is never free.

## The pitch is the outage itself

Sakana did not bury the connection to Anthropic. The release page names the event outright. It cites the export controls imposed on Anthropic's Fable and Mythos models as evidence that access can shift or disappear overnight, and it offers Fugu as the answer: if a single provider restricts access, Fugu dynamically routes around the disruption ([Sakana](https://sakana.ai/fugu-release/)). The whole product is framed as insurance against the thing that had just happened.

Some of the sharpest phrasing around the launch is not actually Sakana's. GIGAZINE, covering the release, wrote that Sakana emphasizes that Fugu allows for flexible model swapping, enabling it to dynamically circumvent the impact even if specific models are restricted ([GIGAZINE](https://gigazine.net/gsc_news/en/20260622-sakana-fugu-multi-agent-system-ai/)). That sentence is worth reading carefully. The "emphasizes that" framing marks it as GIGAZINE's paraphrase, not a verbatim Sakana quote, and the paraphrase is a touch more aggressive than Sakana's own hedged wording. When a claim gets cleaner as it moves from the source to the coverage, that is usually a cue to go back to the source.

The wider trade press read the launch the same way. VentureBeat ran it under the headline "No Claude Fable 5? No problem: Sakana achieves frontier performance with new Fugu multi-model, auto synthesis system" ([VentureBeat](https://venturebeat.com/orchestration/no-claude-fable-5-no-problem-sakana-achieves-frontier-performance-with-new-fugu-multi-model-auto-synthesis-system)), tying the product directly to the outage that set it up.

## Who is building this

Sakana AI was founded in July 2023 in Tokyo by David Ha, its CEO, Llion Jones, its CTO and one of the co-authors of the 2017 paper "Attention Is All You Need" that introduced the transformer, and Ren Ito ([Wikipedia](https://en.wikipedia.org/wiki/Sakana_AI)). The company has a habit of treating existing models as raw material. Its Evolutionary Model Merge work, published in 2024, breeds existing models into new ones ([Wikipedia](https://en.wikipedia.org/wiki/Sakana_AI)), and its AI Scientist project automates parts of the research process itself ([Wikipedia](https://en.wikipedia.org/wiki/Sakana_AI)).

That history is the right lens for Fugu. A company whose earlier splash came from combining models it did not train is a natural author of a product that routes across models it does not own. The idea is continuous with the work, not a swerve away from it.

The funding tracks the ambition. Sakana raised about $200 million in a 2024 Series A backed by Japanese megabanks and Nvidia, added a Series B in November 2025, and reached a valuation near $2.6 billion by late 2025 ([Wikipedia](https://en.wikipedia.org/wiki/Sakana_AI)). For a company barely two years old, that is a large bet on a specific thesis: that the future is many models, coordinated, rather than one model to rule them all.

## What we still cannot confirm

For all the clean framing, several concrete details are not pinned down from primary sources, and it is worth being plain about which ones. The live composition of Fugu's pool is the first. VentureBeat says it draws on OpenAI, Google, Anthropic, and the open-source community in general terms ([VentureBeat](https://venturebeat.com/orchestration/no-claude-fable-5-no-problem-sakana-achieves-frontier-performance-with-new-fugu-multi-model-auto-synthesis-system)), but the specific frontier models in the rotation are not listed on Sakana's own pages, so any precise roster is unverified.

Pricing and benchmark specifics are the second. Tiered prices and per-token Ultra rates have circulated in coverage, and the research said to underpin the coordinator is referenced as well, but none of that was confirmed against Sakana's own pricing or research pages in this review. Where a figure appears only in secondary coverage, this piece leaves it out rather than dress it up as settled.

The legal scope of the original directive is the third. Anthropic's statement stresses that the order targeted foreign nationals, while TechCrunch reports the practical effect was that both models went dark for everyone. Whether US persons were legally required to lose access, or whether Anthropic pulled the models globally as a precaution, is not resolved by the primary statement, and it is the kind of detail that matters if you are trying to reason about how often this can recur.

Two of those open questions can now be pushed a step further. On price, Sakana's own pricing page lists the figures that had only circulated in secondary coverage: subscription tiers of $20 (Standard), $100 (Pro), and $200 (Max) a month, plus pay-as-you-go access to Fugu Ultra at $5 per million input tokens and $30 per million output tokens, rising to $10 and $45 above 272K tokens of context ([Sakana](https://console.sakana.ai/pricing)). Its release page adds that close to 500 early users tried Fugu during a beta before it opened to everyone ([Sakana](https://sakana.ai/fugu-release/)). On the legal question, one analysis argues there is "no clear, comprehensive statutory framework governing government authority to restrict commercial AI model access on national security grounds," that the export rules the Bureau of Industry and Security administers "were not designed with frontier AI models in mind," and that the directive's "specific statutory basis has not been publicly identified with precision" ([Volkov Law](https://blog.volkovlaw.com/2026/06/when-the-government-pulls-the-plug-anthropic-export-controls-and-the-future-of-ai-governance/)).

## The bottom line

The Anthropic shutdown turned a mostly theoretical risk into a documented one. A model you build on can be switched off by a party that is neither you nor its maker, and it can stay off for weeks. Sakana's response is to sell the router instead of the engine: an orchestrator that treats every underlying model as swappable and gives you one endpoint that keeps working when any single provider stops. Whether Fugu's routing is as seamless in practice as the release page suggests is a separate question from whether the problem it names is real. On the evidence of those nineteen days, the problem is real. The rest is a product claim, and product claims are worth checking against primary sources, which in this case are still catching up to the pitch.

## Related reading

- [Planning for a vendor lockout](/blog/blog-ai-vendor-lockout-continuity-plan/): a continuity plan for the exact risk that played out over those nineteen days.
- [AI model routing in 2026](/blog/blog-ai-model-routing-2026/): routing across models as an architecture, the category Fugu sits in.
- [Many small agents beat one big model](/blog/many-small-agents-beat-one-big-model/): the case for orchestrating many models rather than betting everything on one.
- [Multi-model routing in practice](/blog/claude-code-multi-model-routing/): what routing across providers actually looks like day to day.
- [Validate a model before you upgrade](/blog/blog-validate-ai-model-before-upgrade/): why swapping one model for another is never free, and how to check.

## Fact-check notes and sources

- **The shutdown, its basis, and its scope**: the June 12, 2026 directive received at 5:21pm ET, the "narrow, non-universal jailbreak" language, and the statement that other Anthropic models were unaffected come from [Anthropic](https://www.anthropic.com/news/fable-mythos-access); the worldwide practical effect and the "hundreds of millions of people" quote come from [TechCrunch](https://techcrunch.com/2026/06/12/anthropics-safety-warnings-may-have-just-backfired-the-government-has-pulled-the-plug-on-its-most-powerful-ai/) (Connie Loizos); the foreign-national and national-security framing is corroborated by [Al Jazeera](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says). The legal scope for US persons is genuinely unresolved and is presented as such.
- **The lifting**: the June 30, 2026 removal of controls and the nineteen-day framing are from [CNBC](https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html); Anthropic's own confirmation and the restore timing are from its post on [X](https://x.com/AnthropicAI/status/2072106151890809341) and [Al Jazeera](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says).
- **Fugu, what it is and how it is framed**: the June 22, 2026 launch, the taglines, the OpenAI-compatible orchestration, the two tiers, and the footnote that Fable 5 and Mythos are not in the pool are from [Sakana](https://sakana.ai/fugu/) and its [release page](https://sakana.ai/fugu-release/); the "trained a coordinator" distinction and the headline are from [VentureBeat](https://venturebeat.com/orchestration/no-claude-fable-5-no-problem-sakana-achieves-frontier-performance-with-new-fugu-multi-model-auto-synthesis-system); the orchestration-model description is from [MarkTechPost](https://www.marktechpost.com/2026/06/22/sakana-ai-launches-sakana-fugu-an-orchestration-model-that-routes-tasks-across-a-swappable-pool-of-frontier-llms/). The "flexible model swapping" line is [GIGAZINE](https://gigazine.net/gsc_news/en/20260622-sakana-fugu-multi-agent-system-ai/)'s paraphrase, not a Sakana quote.
- **Sakana the company**: founding date, founders, Llion Jones and "Attention Is All You Need," Evolutionary Model Merge, The AI Scientist, and the funding and valuation figures are from [Wikipedia](https://en.wikipedia.org/wiki/Sakana_AI).
- **What is not confirmed**: the exact live model roster, the coordinator's parameter count, and specific pricing and benchmark numbers are not stated on Sakana's primary pages and are omitted here rather than asserted.
- **The terms of the reprieve, and the pushback**: Anthropic's three commitments (proactively detect and address security risks, work with the government on standards for upcoming models, and report any "malicious activity"), the "no longer require an export licence" framing, and the Abraham and Bailo reactions are from [Al Jazeera](https://www.aljazeera.com/economy/2026/7/1/us-lifts-restrictions-on-powerful-ai-models-fable-mythos-anthropic-says).
- **Pricing on the record, and the legal basis**: the $20/$100/$200 monthly tiers and the Fugu Ultra rates ($5 per million input tokens, $30 per million output tokens, rising to $10 and $45 above 272K context) are from Sakana's own [pricing page](https://console.sakana.ai/pricing), and the roughly 500-user beta is from its [release page](https://sakana.ai/fugu-release/); the "no clear, comprehensive statutory framework," "not designed with frontier AI models in mind," and "has not been publicly identified with precision" assessments are from [Volkov Law](https://blog.volkovlaw.com/2026/06/when-the-government-pulls-the-plug-anthropic-export-controls-and-the-future-of-ai-governance/).

*This post is informational and educational, not legal, financial, or investment advice. Company names, product names, and quotations from Sakana AI, Anthropic, and the news outlets cited are reproduced from public sources as nominative fair use, and no affiliation or endorsement is implied. Post-event reporting on the June 2026 shutdown is still developing, and figures not confirmed against primary sources are flagged as such.*


---

Canonical HTML: https://jwatte.com/blog/sakana-fugu-model-router/
RSS: https://jwatte.com/feed.xml
JSON Feed: https://jwatte.com/feed.json
Hero image: https://jwatte.com/images/sakana-fugu-model-router.webp