Mythos Got Banned — Can Asian AI Fill the Gap? What Developers Need to Know About Sakana Fugu and 360's Tulongfeng

Two weeks ago, the US government banned Anthropic's Mythos and Fable 5 models from non-American users. Within days, two Asian companies stepped up claiming they could fill the void. Japan's Sakana AI launched Fugu, and China's 360 unveiled Tulongfeng.

Sounds exciting, right? After spending a day digging through Hacker News discussions and real user experiences, I found the story is way more complicated than the headlines suggest.

The short version: Fugu isn't a model — it's a model broker. Tulongfeng is serious, but only covers security. Neither can truly replace Mythos right now. But this whole episode reveals something bigger: AI geopolitics is here.

What Actually Happened

In mid-June 2026, the US government imposed an export ban on Anthropic, prohibiting Mythos and Fable 5 from being accessible to non-Americans. The justification: these models are "too powerful" and pose a national security risk.

The irony is thick. Anthropic had been pushing for AI regulation for years. Their CEO Dario Amodei repeatedly warned that AI models were "too dangerous" and needed government oversight. Then the government actually listened — and banned their products first. Fable 5 was live for only 72 hours before getting shut down.

The immediate impact hit hard. Developers using Claude Code for daily work, companies building products on Claude's API, entire toolchains dependent on Anthropic's services — all disrupted overnight.

I have friends who were using Claude Code for their daily development workflow. When the ban hit, they were stunned. Going from Fable 5 back to Opus felt like trading a bullet train for a local bus.

The timeline makes it even more interesting. Anthropic had just announced in May that their annualized revenue crossed $47 billion, with a valuation approaching $1 trillion, and they were preparing for an IPO. Then a month later, this happens. Their fastest-growing market segment — Asia — got cut off overnight.

The vacuum was immediate. Developers and enterprises across Asia who relied on Claude's API suddenly had nothing. Who would step in?

Sakana AI's Fugu: Not a Model, But a Model Broker

Sakana AI is a Tokyo-based AI company founded in 2023 by former Google researchers Ren Ito, Llion Jones, and David Ha. They previously raised $135 million in Series B funding at a $2.65 billion valuation. The founders' backgrounds are impressive — Llion Jones is one of the co-authors of the famous "Attention Is All You Need" paper, essentially one of the architects of the Transformer.

On June 22, they released Fugu. The name is interesting — Fugu is Japanese for "pufferfish." Tasty but potentially deadly if not prepared correctly. The name might暗示 something like "powerful but requires careful handling."

Their claim: "Fugu Ultra stands shoulder-to-shoulder with leading models like Anthropic's Fable 5 and Mythos Preview across engineering, scientific, and reasoning benchmarks." Sounds impressive.

But wait. Look at the technical details and you'll spot the issue.

Fugu isn't a standalone large model. It's a model orchestration system.

What does that mean? Fugu itself is a trained language model, but its capability isn't answering questions directly — it's "deciding which models to call to answer questions." It connects to a pool of models — potentially OpenAI's, Anthropic's, or various open-source models — and decides which to use based on task complexity.

In their own words: "Rather than a single monolithic model, Fugu is a learned multi-agent orchestration system."

This raises an obvious question. If Fugu is calling Anthropic's models behind the scenes, wouldn't the ban affect it too? Sakana's answer is "the model pool is swappable," but that's a bit like saying "our system is strong because we can swap engines" — swap the engine, and is it still the same car?

What Is an Orchestration Model? Why Should You Care?

Before going further, I think it's worth explaining the "orchestration model" concept, because this might be an important direction for AI's future.

Traditional AI models are "one model does everything." GPT-5.5, Claude Opus, DeepSeek V4 — whatever you ask, the same model answers. It's like having a generalist employee who's decent at everything but not necessarily great at any one thing.

Orchestration models work differently. They're more like a project manager with a team of specialists — some excel at coding, some at math, some at writing — and the manager assigns work based on the task. You don't need to know who's doing the work; the manager handles it.

Sakana Fugu follows this approach. It's a trained language model, but its capability is "deciding which models to call" rather than directly answering questions. It can decide: simple tasks it handles itself, complex tasks get delegated to more capable models.

The academic foundation comes from two papers Sakana AI published at ICLR 2026: Trinity and Conductor. The core idea is "learning to coordinate" rather than "learning to answer."

In theory, this is elegant. In practice?

Pros:

Can leverage each model's strengths while avoiding weaknesses
Model pool is swappable, avoiding single-vendor lock-in
Potentially more effective for complex multi-step reasoning tasks

Cons:

Cost stacking — calling multiple models means paying multiple times
Latency increase — each decision takes time
You don't know which model is actually doing the work, low transparency
If the underlying models get banned, the orchestration is useless

Orchestration is essentially a "leverage" strategy. It doesn't build its own wheels; it learns to use other people's wheels. This works great when there's a rich ecosystem of models available, but if those wheels get taken away, leverage can't move anything.

Real User Feedback from Hacker News: Don't Get Too Excited

The Hacker News thread about this news got 238 points and 178 comments. The vibe in the comments was very real.

One person said they tried Fugu, and a single prompt exhausted their $20 plan's 5-hour quota. Upgrading to the $100 plan, they found the results were "worse than Opus, incredibly slow." Their exact words: "the result was worse than Opus, incredibly slow, and I ended up exhausting the new 5 hour window and have used 35% of the weekly now."

Another person's explanation was insightful: "If fugu really is an orchestrator dispatching to opus/gpt under the hood, the $20-in-one-prompt complaints actually start making sense — you're paying api markup twice."

Meaning: if Fugu is calling Opus or GPT APIs behind the scenes, you're paying Fugu's API fee + the underlying model's API fee. Double middleman markup.

However, some people had positive experiences. A security engineer said Fugu Ultra was excellent for security assessments: "Fugu drove a full security assessment end-to-end — recon, XSS/SQLi checks, auth review, and a clean report with evidence and retest steps."

Overall sentiment: people are very skeptical of the "Mythos-level" claim.

"Without reliable benchmarks, they are Mythos-like only in the sense that they accept text as input and produce text as output." This comment got a lot of upvotes.

Fugu's Pricing and Real-World Experience: Is Your Wallet Ready?

Let's talk practical. Fugu's pricing structure:

Fugu (standard): Suitable for daily use, lower latency
Fugu Ultra: For complex tasks, calls more models, higher quality but more expensive

From HN user feedback, pricing is a major concern.

One developer used Fugu for a Unity project code review. Their $20 plan's 5-hour quota was exhausted on a single prompt. After upgrading to the $100 plan, they found the results were worse than Opus, and incredibly slow. The 5-hour quota was used up again, consuming 35% of their weekly limit.

"the result was worse than Opus, incredibly slow, and I ended up exhausting the new 5 hour window and have used 35% of the weekly now."

Another person's explanation was spot on: "If fugu really is an orchestrator dispatching to opus/gpt under the hood, the $20-in-one-prompt complaints actually start making sense — you're paying api markup twice."

Meaning: if Fugu calls Opus or GPT APIs behind the scenes, your cost = Fugu's API fee + underlying model's API fee. Double middleman.

However, some found it worth it. A security engineer said Fugu Ultra was great for security assessments: "Fugu drove a full security assessment end-to-end — recon, XSS/SQLi checks, auth review, and a clean report with evidence and retest steps."

My verdict: Fugu might be better suited for enterprise-level complex tasks, not ideal for individual developers' daily use. If you're just writing code and fixing bugs, directly calling Opus or GPT-5.5's API might be more cost-effective.

Fable 5's Real Experience: What Users Said Before the Ban

An interesting side topic: what did people who actually used Fable 5 before the ban think?

HN discussions provided some firsthand accounts.

Someone used Fable 5 with Claude Code CLI for a full day of development and was blown away: "It acted like a senior engineer - actually coding up hypotheses, testing them, finding problems and presenting good, usable recommendations backed by solid evidence and wisdom. It can probably do most of my job, which gave me a bit of an existential crisis."

But others had mediocre experiences. Someone tested Fable 5 in Cursor for CSS styling, and it "spun out the most useless, Claude-like CSS styling ever, wasting $40 in 10 minutes."

One person made an interesting comparison: Fable 5 vs Opus 4.8 on large legacy code modernization. Their conclusion was that Fable 5 was indeed stronger, but not overwhelmingly so. "It's more like an optimization, I could have a single or 2 pass in fable vs 8-10 with opus to arrive at the same solution."

This is important context. It shows Mythos/Fable 5's power isn't magic — it's an "efficiency boost." If Opus takes 10 attempts to achieve what Fable 5 does in 2, that's significant. But for simple tasks, the gap isn't that large.

So when Sakana claims Fugu "matches Fable 5," consider: if Fable 5 itself is essentially "Opus optimized" for many tasks, how much is Fugu's "matching" really worth?

360's Tulongfeng: This Time It's Serious

If Sakana Fugu has a hint of "riding the hype," 360's Tulongfeng is a completely different story.

360 is a veteran Chinese cybersecurity company. Their founder Zhou Hongyi was direct: vulnerability-finding AI is a national strategic asset. Their two products — Tulongfeng (automatic software vulnerability discovery) and Yitianzhen (automated cyber defense and incident response) — directly target Anthropic Mythos's capabilities in the security domain.

And 360 didn't hide behind "coincidental timing." They explicitly mentioned the risk of "one-way transparency" — a situation where some actors can access advanced vulnerability detection while others cannot.

Honestly, from a technical perspective, 360's positioning is more pragmatic than Sakana's. They're not claiming to "surpass Mythos across the board" — they're saying "in the security vertical, we need our own capabilities." This approach is more grounded.

Background: The Big Picture of AI Export Controls

To understand the full picture, you need to see AI export controls in a broader context.

US government restrictions on AI models aren't new. Earlier restrictions targeted high-end GPUs (like NVIDIA H100), banning their sale to China. Now the restrictions have expanded to models themselves — not just hardware, but software capabilities can't be freely exported.

The underlying logic: AI capabilities are becoming national security assets. An AI that can automatically discover software vulnerabilities might be more valuable in cyberwarfare than a missile. So governments want to control the spread of such capabilities.

But here's the thing: technology diffusion isn't as easy to control as physical products. Ban Anthropic's API? People can train their own models with open-source alternatives. Ban GPUs? People can source computing power through other channels. Technology, once invented, is hard to completely contain.

That's why Asian companies responded so quickly to the Mythos ban. Not because they happened to be ready, but because the market opportunity was too big — someone would always step in to fill the gap.

Anthropic's Awkward Position

This might be the most ironic part of the whole story.

Anthropic had been pushing for AI regulation for years. Their CEO Dario Amodei repeatedly stated that AI models were "too dangerous" and needed government intervention. Then the government actually intervened — and banned their products first.

Someone on HN commented sharply: "Anthropic are the pathetic ones. The pariah of the AI industry that nobody likes because all they do is lie, cheat and steal. Now no one can access ChatGPT 5.6 because of their 5 year long fearmongering regulatory capture campaign."

That's a bit harsh, but it reflects a sentiment: many people feel Anthropic's "safety advocacy" ultimately shot themselves in the foot. You kept saying "AI is too dangerous," the government listened, and then banned your product. The script is almost too ironic.

Though from another angle, the ban might be intentional. By pushing for regulation, Anthropic can build higher barriers to entry — only "responsible" companies like themselves can develop and deploy advanced AI. This strategy hurts short-term but might build a moat long-term.

Regardless, the current situation is that Anthropic's revenue might be significantly impacted. They just announced in May that annualized revenue crossed $47 billion, and now their biggest growth engine — the Asian market — has been cut off. Not great news for their upcoming IPO.

User Experience: Extreme Polarization

Regarding Fable 5 and Mythos user experiences, HN discussions showed clear polarization.

Positive reviews said Fable 5 was like a senior engineer, capable of independently completing complex tasks. Some said it excelled at security assessments, able to complete reconnaissance, vulnerability detection, authentication review, and report generation end-to-end.

Negative reviews said it wasted money, was slow, and produced results worse than Opus. Some said the code style it generated was identical to Claude's — no breakthrough whatsoever.

This kind of polarized experience is actually common. AI model performance is highly dependent on task type and usage patterns. Direct CLI calls might work great, but calls through intermediary layers like Cursor might lose quality. Same model, different prompt engineering, vastly different results.

So when you see someone say "this model is amazing" or "this model is garbage," take both with a grain of salt. The only way to know is to try it yourself.

The Competitive Landscape: Not Just Japan and China

It's not just Sakana and 360 making moves. The entire Asian AI landscape is shifting.

DeepSeek already proved that China can produce high-quality open-source models at lower cost. GLM-5.2 is also performing well on open-source leaderboards. Japan has Sakana, South Korea has Naver's HyperCLOVA, and India is developing its own sovereign AI.

The Mythos ban has somewhat accelerated this trend. Previously, people could get by with "American models are the best, just use American ones." That doesn't work anymore — when access can be cut at any moment, not having alternatives means being completely exposed.

Sakana's co-founder Ren Ito said at the G7 summit: "AI should not become a technology that is hoarded; it should be one that is developed together."

Ideals are nice. Reality is: technology competition is intensifying, and every country is developing its own AI capabilities to avoid being strangled by others.

Developer Perspective: What Does This Mean for You?

Enough background. Let's talk about what directly matters for developers.

First, Don't Blindly Trust "Mythos-Level" Claims

More and more companies are claiming their models are "Mythos-level." But here's the problem: most developers have never used Mythos, so how do you judge?

Someone on HN put it well: "I don't even look at benchmarks anymore. I just try different models on our large, proprietary, systems software codebases in real, shipping products."

Benchmarks can be gamed. Marketing can be hyped. But whether it actually works well in practice, only you know. If you're considering switching models, don't just look at claims — run it on your real tasks.

Here's how to evaluate practically:

Test on your real codebase, not hello world or toy examples
Run at least 5 different task types: code generation, bug fixing, code review, documentation, architecture discussions
Record results: how much money spent, how much time used, quality assessment
Compare with your current solution: if a new model is 20% better but 3x more expensive, it might not be worth switching

Second, Orchestration Models Are a Promising Direction, But Not a Silver Bullet

Sakana Fugu's "orchestration model" approach is genuinely interesting. Having one model decide which models to use could theoretically combine strengths while avoiding weaknesses. But there's a big practical issue: cost and latency.

One user reported that a single Fugu code review prompt exhausted their $20 plan. If it's calling multiple model APIs behind the scenes, costs do stack up. For individual developers, this pricing might not be friendly.

For enterprise users, though, if Fugu genuinely reduces manual review workload, the cost might be worthwhile. The key is doing the math carefully.

Third, AI Sovereignty Is Real

Sakana's co-founder David Ha said something perceptive: "Access to top models can disappear overnight."

This isn't hypothetical — it already happened. The Mythos ban proved one thing: your AI capabilities can be cut by a stroke of a pen. For companies that depend on AI APIs for their business, this is a tangible risk.

For individual developers, this means:

Don't rely on a single model provider. Have at least two: one primary, one backup
Follow open-source model progress. DeepSeek, Qwen, Llama won't get banned
Consider local deployment possibilities. Local models aren't as capable as cloud ones, but they're sufficient for many tasks

Fourth, Open-Source Models Are the Last Line of Defense

Export bans hit closed-source models hardest. If you're using open-source models — DeepSeek, Qwen, Llama — export bans barely affect you. You can run them on your own servers without depending on any company's API.

This is why open-source model interest has been surging. Not because open-source models are stronger than closed-source ones (mostly they're not yet), but because open-source models can't be banned.

And open-source models are advancing fast. DeepSeek DSpark recently open-sourced their inference acceleration solution, boosting inference speed by 85%. GLM-5.2 is also performing well on open-source leaderboards. The gap is narrowing.

Fifth, Cost Control Matters More Than Model Choice

Many people agonize over "which model to use," but "how to use it" is more critical.

Regardless of which model you use, if you throw your entire codebase into the context every time, costs will be high. Learning to control context length, using RAG for retrieval augmentation, breaking large tasks into smaller ones — these techniques save more money than model selection.

I've discussed headroom before, a tool that can save 60-95% on context overhead. Whether you're using Fugu or Opus, learning to control costs is essential.

My Verdict

Honestly, can Sakana Fugu and 360's Tulongfeng fill the Mythos gap? Short-term: probably not.

Fugu is fundamentally an orchestration system — its strength depends on the underlying models it calls. If those models get banned too, Fugu becomes useless. And from user feedback, cost and experience issues are still significant.

360's Tulongfeng is more practical — at least it has genuine expertise in the security vertical. But claiming it fully matches Mythos? I think that's premature.

However, the emergence of these companies is itself a signal: AI's landscape is shifting from "one dominant player" to "multi-polar competition."

Anthropic got banned, Sakana stepped up. America restricted, Chinese and Japanese companies filled in. This trend won't stop. For developers, the good news is you have more choices. The bad news is more choices mean harder decisions.

My advice: don't rush to switch. Wait and watch. Let these new models run for two or three months, accumulate real user feedback and third-party evaluations, then make your decision. Jumping in now likely means being a guinea pig.

If you need an alternative right now, here's my priority list:

Try Opus 4.8 first — not Mythos-level, but currently the strongest available closed-source model outside the ban
DeepSeek V4 + Claude Code — open-source model with coding tools, great cost-efficiency
Wait for Fugu's third-party reviews — don't be an early adopter
Watch 360's product releases — if you do security-related development

I'm planning to experiment with Fugu's API myself, running several real tasks to see actual results. If you've already tried Fugu or 360's products, feel free to share your experience in the comments.

Written on June 28, 2026, based on TechCrunch reporting, Sakana AI's official announcement, and Hacker News community discussions. Data accurate as of publication.*