Fugu Ultra vs Mythos and Fable: how good is Sakana’s new AI orchestrator really?

When Anthropic’s top models Mythos and Fable 5 were suddenly pulled from public access due to US export controls, it exposed a big risk for anyone building serious systems on a single AI provider. Just days later, Japanese lab Sakana stepped into that gap with Fugu, an AI "orchestrator" that claims to match the performance of those now-restricted models—without being tied to one vendor.

What Fugu actually is (and isn’t)

Fugu isn’t a giant frontier model trying to outmuscle GPT or Claude. Instead, it’s a relatively small model—around 7 billion parameters—whose main job is to decide which other model should handle each part of your request.

You send one prompt to Fugu. Behind the scenes, it:

• Picks which underlying models to use
• Splits your task into pieces
• Sends each piece to the model best suited for it
• Checks and combines the answers
• Returns a single, unified response

Sakana calls this orchestration. A helpful analogy is a general contractor: Fugu doesn’t do all the work itself, but it knows which specialist to call for each job—coding, reasoning, 3D graphics, research, and more. It can even call itself again as part of a larger workflow.

Who is behind Sakana and Fugu?

Sakana isn’t a random new startup. Two of the key people behind it are well-known in the AI world:

• David Ha, CEO, previously at Google Brain
• Llion Jones, co-founder, one of the authors of the landmark "Attention Is All You Need" paper that kicked off the Transformer era

That background matters because Sakana emphasizes that Fugu’s routing is actually trained, not just a pile of hard-coded if-statements. The approach is backed by two research papers accepted at a conference this year, positioning Fugu as a serious research-based system rather than a quick wrapper.

Fugu vs Fugu Ultra: two different modes

Fugu comes in two main versions:

Fugu (standard)
The faster, everyday option. It aims for good performance at lower cost and latency, suitable for typical chat, coding, and productivity tasks.

Fugu Ultra
The heavyweight mode. It pulls in deeper and more capable models for harder work, including:

• Technical research
• Security analysis
• Patent search
• Complex simulations and 3D rendering

Most of the eye-catching examples and benchmark claims you’ll see online refer specifically to Fugu Ultra.

Benchmark claims: can Fugu really match Mythos and Fable?

On Sakana’s own charts, Fugu Ultra posts very strong numbers:

• 93.2 on LiveCodeBench (coding benchmark)
• 95.5 on GPQA Diamond (hard science reasoning test)

On paper, these scores allow Fugu Ultra to edge out some top models like Claude Opus 4.8 in certain benchmarks, and Sakana suggests it stands “shoulder to shoulder” with Anthropic’s Mythos and Fable models.

However, there are two big caveats:

• All numbers are self-reported by Sakana—there’s no broad independent verification yet.
• Comparisons are based on published benchmark scores, not live head-to-head tests against Mythos or Fable.

If you’ve been following the Mythos/Fable story and how they’re expected to reshape security and reasoning, you may want to read deeper context in this breakdown of Anthropic’s Mythos line.

Why Sakana is pushing “AI sovereignty”

Sakana’s bigger argument isn’t just “our numbers are high.” It’s about infrastructure risk.

When a single provider’s model underpins critical systems—banks, governments, infrastructure—export controls or policy changes can cut off access overnight. That’s exactly what happened when Mythos and Fable 5 disappeared from public use.

Fugu’s answer is redundancy:

• Don’t rely on one model or one company
• Use an orchestrator that can route around outages, bans, or pricing changes
• If one provider is cut off, the orchestrator silently shifts the workload elsewhere

Sakana frames this as "AI sovereignty": owning your AI stack at the orchestration layer so you’re not at the mercy of any single vendor’s decisions or geopolitics.

Pricing and access: how much does Fugu cost?

For Fugu Ultra, Sakana’s pricing is:

• $5 per million input tokens
• $30 per million output tokens

There are also subscription plans ranging from roughly $20 to $200 per month, depending on usage tiers.

Two important limitations:

• Blocked in the EU: Fugu is currently unavailable in the European Union due to privacy regulations.
• Black-box routing: You can’t see exactly which underlying models Fugu used for your request. That’s intentionally hidden, which makes it feel like a black box from a transparency and compliance perspective.

Real-world tests: Fugu Ultra vs Opus and GPT

Early community tests paint a more mixed picture than the benchmarks.

Coding a simple game

In one comparison, both Claude Opus 4.8 and Fugu Ultra were asked to build a classic “chicken crossing the road” style game:

• Fugu Ultra finished in about 22 minutes
• Opus 4.8 took around 79 minutes
• Fugu’s cost: ~$7.32
• Opus 4.8’s cost: ~$37.85

Visually, the Opus result was arguably more polished, but both outputs looked comparable at a glance. If you care more about speed and cost than aesthetics, Fugu Ultra clearly wins this particular test. But without deep playtesting, it’s hard to say which implementation is more robust.

3D and 3.js generations

Where Fugu Ultra really stands out is 3D-related work. In several 3.js and simulation prompts, Fugu Ultra produced results that looked noticeably better than single models like GPT 5.5 or Opus 4.8 when used alone.

This makes sense: 3D rendering and complex visual simulations often require multiple skills—math, graphics programming, physics, and code structuring. An orchestrator that can combine strengths from multiple models has a natural advantage here.

Some users have even compared Fugu Ultra’s 3D outputs to the kind of rich simulations people saw from Fable before it was pulled, though side-by-side, Fable still appears ahead in quality.

How Fugu compares to GLM 5.2 and other non-US models

Fugu isn’t just competing with US frontier models. It’s also being compared against top Chinese models like GLM 5.2 and MiniMax.

Key observations from early tests:

• GLM 5.2 is very competitive with GPT 5.5 and Opus 4.8 on many prompts.
• Fugu Ultra often produces the best-looking or most capable output in 3D and simulation tasks.
• However, GLM 5.2 is significantly cheaper—around $0.03 in one example, versus $0.51 for Fugu on the same prompt.

In some runs, Fugu Ultra is actually more expensive than Opus 4.8 or GPT 5.5, depending on which underlying models it decides to call. That means you can’t assume Fugu is always the cheapest or the best—it’s highly prompt-dependent.

If you’re interested in how non-US models are starting to challenge US leaders more broadly, you may also find this comparison of local and cloud models useful: can a 35B local model really beat Claude Sonnet 3.5?

Is Fugu really at Mythos/Fable level?

Despite the strong benchmarks, many hands-on testers are skeptical of the claim that Fugu Ultra truly rivals Anthropic’s Mythos and Fable 5 in practice.

Some recurring feedback:

• Reliability is inconsistent—sometimes Fugu takes too long or feels unstable.
• Quality varies by prompt; it can shine in 3D and simulations but feel merely “on par” in more standard tasks.
• Side-by-side comparisons with known Fable generations (like complex “My Little Pony” style simulations) show Fable still clearly ahead in richness and coherence.

So while Fugu may beat top models on certain benchmarks and specific tasks, the broader claim that it fully matches Fable 5 and Mythos 5 doesn’t seem to hold up yet in real-world usage.

Where Fugu makes sense—and where it doesn’t

Fugu’s real value isn’t that it’s a magical new model. It’s that it’s a smart orchestrator sitting on top of multiple strong models.

Fugu is most compelling if:

• You want resilience against model bans, outages, or policy shifts.
• Your workloads benefit from combining skills across models (e.g., complex coding + 3D + reasoning).
• You’re okay with a black-box routing layer in exchange for convenience and performance.

It’s less compelling if:

• You mainly care about predictable pricing and transparency.
• A single strong model like GLM 5.2, Opus 4.8, or GPT 5.5 already covers your needs.
• You need strict compliance, auditability, or full visibility into which models are being used.

In that sense, Fugu feels more like an infrastructure play than a pure model breakthrough. It shows how powerful model orchestration can be—but it also highlights the trade-offs in cost, transparency, and reliability that come with that approach.

Bottom line

Fugu Ultra is an ambitious attempt to answer the “what if my AI provider disappears?” question that Mythos and Fable’s removal made very real. It can be faster and cheaper than individual frontier models on some tasks, and it’s especially impressive for 3D rendering and simulations.

But the hype that it fully replaces Fable 5 or Mythos 5 is, so far, ahead of reality. Benchmarks look great, yet real-world results are mixed, costs can be unpredictable, and the black-box nature of its routing won’t suit everyone.

If you’re building serious AI systems, Fugu is worth watching as a sign of where orchestration and AI sovereignty are headed—even if it’s not yet the definitive answer to losing access to top-tier models.