Journal
Multi-model orchestration has been the talk of the last few weeks. The example everyone’s citing is Sakana Fugu, the router from a Tokyo lab that farms requests out across GPT, Gemini, and Anthropic’s models. A lot of the coverage framed it as the first crack in US dominance over frontier AI, talk of a power shift east, of the American lead cracking. Even Sakana walked that framing back within days: a spokesperson described it as a moment, not a permanent realignment toward a different set of players. And the loudest early reaction, on Hacker News and elsewhere, wasn’t geopolitical triumphalism at all, it was technical skepticism, a lot of people asking whether Fugu was actually new or just an expensive router wearing a model’s clothes.
What actually matters isn’t who tops the leaderboard. It’s that Sakana’s own CEO said the quiet part out loud: orchestration models are the next frontier, beyond bigger models.
He's right. Just not entirely for the reason he means.
The Shift
Value is moving from the model that answers to whoever decides which model should answer, and with what instructions. That’s not a discovery from ten days ago. LangChain and CrewAI have been hand-wiring model orchestration for a while, OpenRouter does multi-model synthesis with Fusion. What changes with a system like Fugu is that the coordination itself is learned rather than hand-coded, a coordinator trained to decide who to call, in what order, when to verify the work. Fugu in particular exists partly as a hedge against a very concrete risk: two of the models it benchmarks against, Anthropic’s Fable 5 and Mythos Preview, were pulled from its own pool on June 12, 2026, when they became subject to US export controls. But the underlying mechanism, value moving from the model to the layer that uses it, was already underway before this one launch.
The Hole in the Argument
An orchestrator, by definition, routes. And it routes based on whatever context it’s given: brand voice, rules for what to say and avoid, what’s already been said last week. If that context is vague or doesn’t exist in structured form, routing doesn’t fix it. It just executes it faster.
Picture an editor-in-chief assigning stories to genuinely talented freelancers. Without a clear editorial line, who the readers are, what already ran, the assigning still happens. Every piece comes back well written and disconnected from the rest. The team didn’t get worse. It’s working without the material judgment requires.
The early real-world tests on Fugu fit this frame too, even though they’re pointing at a slightly different problem. Ethan Mollick wrote that Fugu Ultra was painfully slow on his usual coding tests, roughly thirty minutes a run, with results he called fine but not on par with Fable in real use. None of this is an argument against Fugu. It’s a reminder that a routing engine, however sophisticated, doesn’t compensate for thin context upstream. At best it processes that thinness more efficiently.
The Undefended Layer
That data context matters is widely accepted in enterprise circles by now, semantic layers, ontologies, data governance are treated as serious infrastructure, with analysts openly saying the model is commodity and context is the real moat. What’s missing almost everywhere is that same treatment applied to brand context. Voice, audience, editorial guardrails, what’s already been published, in most organizations this lives in a shared doc nobody updates. It isn’t structured, it doesn’t compound over time, and it isn’t queryable by a system deciding in real time who gets routed what.
What Stays
With Fylle we own exactly that layer: context optimized for AI and readable for people. A system that compounds over time and stays owned by whoever uses it, always, not sitting in some vendor’s temporary cache.
Execution is commodity now. Models change. But the context, the memory, what the AI knows about you and what you want it to know and remember, that stays.
For the companies we work with, a model-agnostic system is the right answer: it works across OpenAI, Anthropic, Gemini, DeepSeek, which opens up two things that aren’t a footnote for certain industries:
- Open-weight models, not just closed proprietary APIs.
- On-prem deployment, the real constraint for anyone in regulated industries like finance or pharma, where data can’t leave for a third-party managed cloud orchestrator.
And in the same week we learned access to a model can vanish over an export control order signed by a government, that stops being a positioning detail.
Sources
- Sakana AI, Fugu release announcement, June 22, 2026: sakana.ai/fugu-release
- Ethan Mollick on Fugu Ultra in real-world testing, as reported by Business Standard, June 23, 2026: business-standard.com
- TechCrunch, on Sakana’s positioning and spokesperson comments, June 27, 2026: techcrunch.com
- Anthropic, export control announcement on Fable 5 and Mythos, June 12, 2026: anthropic.com/news