Here is the uncomfortable bit. For two years the winning move was simple. Pick the biggest, smartest model, type your question, take the answer. One coin in, one snack out. That instinct is now the weakest way to use AI, and a small Tokyo lab just built the proof.
Angus Sewell (Angus the Nontechnical) argues that the edge has moved from owning the cleverest model to orchestrating a room full of them, and he points at Sakana’s new system, Fugu, as the turning point. He is right, and the mechanism is worth understanding in plain terms. Orchestration just means a manager model that reads your task, hires one model to plan, another to do the work, a third to check it, then blends the results into a single answer. It manages the specialists rather than trying to be one.
Sakana is the lab co-founded by one of the authors of the original Transformer paper, the research that kicked off this whole circus, so they are not cranks. Their top variant, Fugu Ultra, scored 73.7 on SWE-Bench Pro and 95.5 on GPQA-D, and on Sakana’s own tests it leads 10 of 11 benchmarks, sitting ahead of Opus 4.8, Gemini 3.1 Pro and GPT-5.5. Co-founder David Ha put the thesis bluntly: “Orchestration Models are the next frontier, beyond bigger models.”
Now the pushback, because you should always squint when a lab grades its own homework. Those headline numbers come from Sakana’s published benchmarks, not from an independent referee, and “we beat everyone on the tests we chose” is the oldest move in the book. Treat the leaderboard bragging as marketing until someone neutral repeats it.
But here is the part that actually matters for you, the person paying the invoice, and it is not the leaderboard at all. Independent coverage of Fugu notes it runs a large share of queries, often the majority, on the cheap tier and still lands near top-model quality, because the router predicts difficulty up front and only calls the expensive model when it has to. That is the real story. Not a genius in a box, but a way to stop paying frontier prices for questions a budget model could answer. For a solo operator or a small team, smart routing is where the money leaks back into your pocket.
So do not rush to rip out your setup. The lesson is cheaper and calmer than the hype. Stop treating one model as an oracle, start treating your stack as a team, and let something route the easy 80 per cent to the cheap seats.
Did you know: the router adds only milliseconds of delay and no extra inference cost, because it decides where to send your query before any big model even wakes up.