Central Casting

The Model Is the Talent. Fugu Is the Agent.

A Tokyo lab called Sakana shipped something Monday that should put a knot in every frontier CEO's stomach, and most people filed it under "neat demo." It's called Fugu, and it doesn't compete with GPT-5.5 or Claude Opus or Gemini. It bosses them. You send one request. Fugu decides which model is best for which slice of the job, farms the slices out, checks the work, and hands back a single answer. On SWE-Bench Pro it scored 73.7 and beat every model it was managing — Opus 4.8 at 69.2, GPT-5.5 at 58.6. It never trained a frontier model of its own. It just learned which one to call.

Sit with what that does to the question we've all been asking. For two years the only thing that mattered was who had the smartest model. That's the wrong question now. The new one is who decides where the smarts get spent, and that's a different company entirely.

The cleanest way to see it: Fugu is DNS for intelligence. Most of you never think about DNS — the invisible switchboard that turns a name into an address — and you ride whatever your provider hands you. The people who know better point at Cloudflare or Google and get something faster. But a DNS resolver only changes how fast you reach an answer. The router changes the answer itself. Pick the wrong resolver, you lose milliseconds. Pick the wrong model, you ship a broken migration or a hallucinated contract.

We've been saying the model is a rental till it's boring. The part that isn't: the rental is now priced by the hour, in a live market, by an agent that doesn't care whose name is on the engine. Crack a hard security problem? Expensive tier. Score a stack of articles? A cheap open model on your own box does it for nothing. The router skims the spread, and the spread is wide — the same job has run six times dearer on a premium model than a discount one.

So the field flattens to a barbell. Be the model on the absolute frontier, or be the cheapest one that gets the average job done. Everything in the soft middle gets routed around and compressed to zero. The middle isn't a place you sit — it's a slope you slide down, and Monday's frontier is next quarter's bench player.

Every model now auditions for every job, and the only seat that matters holds the casting sheet. Own it, or let a vendor's default decide who you hire.

At COAI today: the full Signal/Noise — the routing-table flywheel, why this is an enterprise story, and the eight questions to ask before you standardize on anybody's model — is live at getcoai.com.

Outsider Labs Advisory →

73.7 — Sakana Fugu's score on SWE-Bench Pro, beating Claude Opus 4.8 (69.2) and GPT-5.5 (58.6) without training a single model of its own. It's not a model. It's a conductor that hires the models, checks their work, and ships one answer better than any of them alone. The best model in the room just became a freelancer waiting for a callback — and the company that decides the callback list never wrote a line of frontier code.

Stop picking a house model. Start building a routing table. Standardizing the whole company on one frontier vendor is the bet that ages worst — the leader changes every few weeks. Today: list your top five AI workflows and name, per workflow, the cheapest model that clears the bar. That table is now a real asset. Start keeping it.

Route by the task, not by the leaderboard. SWE-Bench tells you who codes well. It tells you nothing about who writes a board memo your CFO won't redline, or which image model fits your brand. The valuable routing calls are the ones with no public benchmark. Today: pick one fuzzy, high-volume task — support replies, first-draft copy — and run it across three models on the same prompt. Trust your eyes, not the chart.

Treat American-model access as revocable. Washington pulled a frontier model off the global market this month on ninety minutes' notice. Fugu's whole pitch is that it routes around any provider that goes dark. Today: make sure no critical workflow has exactly one model behind it. Single-vendor intelligence is now concentration risk, and your auditor will eventually call it that.

Tuesday, June 23

Sakana's Fugu makes every frontier model a temp worker. (Full analysis above.) A Tokyo orchestrator routes one API call across GPT, Claude, and Gemini and beats all of them on SWE-Bench Pro. It excludes the EU pending GDPR and routes around export controls by design — the model brands vanish into it like interchangeable organs.

Nadella opens a price war on his own partners. In a Sunday WSJ interview Microsoft's CEO pushed for cheaper, swappable models and "social permission." Read the balance sheet: Azure runs ~$37B a year mostly on other people's models, so driving inference to zero is free for the landlord and lethal for the tenants — OpenAI and Anthropic.

Bezos says the bulldozer elevates, it doesn't bury. On CNBC he called the no-more-radiologists, no-more-engineers crowd flat wrong — AI lifts the worker an altitude, and the productivity surge brings a labor shortage and deflation, not a jobless apocalypse. The most credible builder alive just took the other side of the jobs trade.

An AI law firm won a real trial for about £400. Garfield AI, the UK's first regulated AI firm, did all the pre-trial work in a case revealed Monday; the claimant recovered £7,000 against a defendant who paid for a solicitor and a barrister. The advocacy stayed human. The legible 80% didn't.

China closed the gap to seven months — without Nvidia. GLM 5.2 narrowed the frontier gap to roughly seven months, DeepSeek is now valued at $7.4B, and Huawei's CloudMatrix post-trained a 1.6-trillion-parameter model on zero Nvidia silicon. The "they can't catch up" story needs a rewrite.

— Harry and Anthony

Sources: