Sakana AI announced Fugu and Fugu Ultra on June 22. The system is a multi-agent orchestration platform, not a single model. Multiple specialized AI agents coordinate behind a single API. To the user, it looks like one model.
The system dynamically routes tasks to the most capable sub-agent for each job. For complex tasks, it can spin up specialized agent teams on the fly. The user never needs to choose a model or manage multiple API keys.
Fugu balances strong performance with low latency for everyday work like coding and chat. Fugu Ultra is optimized for research, paper reproduction, cybersecurity analysis, and patent investigation.
The Benchmark Claims
Sakana AI claims Fugu Ultra matches or exceeds Fable 5 and Mythos Preview on several benchmarks. The data shows:
Benchmark | Fugu Ultra | Fable 5 | Mythos Preview | GPT-5.5 |
|---|---|---|---|---|
Terminal-Bench 2.1 | 82.1 | 74.6 | 70.3 | 78.2 |
CharXiv Reasoning | 86.6 | 84.2 | 83.3 | 84.1 |
SWE Bench Pro | 73.7 | 69.2 | 54.2 | 58.6 |
LiveCodeBench Pro | 90.8 | 84.8 | 82.9 | 88.4 |
Humanity‘s Last Exam | 50.0 | 49.8 | 44.4 | 41.4 |
Fugu Ultra outperforms on Terminal-Bench 2.1 and CharXiv, while trailing slightly on Humanity’s Last Exam. Overall, Sakana describes it as“shoulder-to-shoulder” with frontier models.
The caveat: Fable 5 and Mythos Preview are not publicly accessible, so comparisons are based on Sakana‘s own testing. The claims are credible but not independently verified.
The Export Control Angle
This is the part that matters. Sakana AI explicitly addressed the Fable 5 shutdown in its announcement. The company noted that regulatory frameworks and export controls can alter access overnight — referencing the June 12 US order that cut off non-US access to Anthropic’s models.
The company said relying on a single API for critical infrastructure is“a real weakness” for organizations and nations. Fugu runs on a pool of entirely swappable agents. If one provider restricts access, Fugu routes around the disruption.
This is not a technical claim. It‘s a product positioning claim. Sakana AI is selling resilience, not performance. The model is the hedge.
Pricing
Fugu is available through subscriptions and pay-as-you-go:
Subscription plans:
Standard: $20/month
Pro: $100/month
Max: $200/month
Fugu Ultra pay-as-you-go: $5 per million input tokens and $30 per million output tokens. For contexts over 272K tokens, prices increase to $10 input and $45 output.
What This Actually Means
Sakana Fugu is not a breakthrough. It‘s a workaround. OpenAI and Anthropic sell models. Sakana AI sells access to a team of models that can be swapped if one gets blocked. The strategy is simple: aggregate other people’s models, optimize task routing, and sell the result as a single API. It competes with Fable 5 on performance by combining models that are weaker individually but stronger together.
The timing is strategic. Fable 5 users are still waiting for the model to return from regulatory limbo. Sakana AI is offering an alternative that works, costs less, and doesn‘t face the same compliance uncertainty. The model race isn’t the only race anymore. The orchestration race just started.
P.S. Fugu Ultra doesn‘t beat Fable 5 on every benchmark. It beats it on some benchmarks. But the benchmark isn’t the point. The point is that Fable 5 is offline. Fugu is not. Availability beats performance when the product you need isn‘t on the shelf. Japan’s AI strategy appears to be: don‘t build the best model. Build the system that survives when the best model disappears.
