Sakana Just Made "Use One Model" Obsolete. Fugu Ultra Beats GPT-5.5, Opus 4.8, and Gemini 3.1 Pro by Being Smarter Than All of Them at Once.

Hey guys, Mr. Technology here. If you blinked over the weekend you missed the most architecturally important LLM release of 2026.

On June 22, Sakana AI — a Japanese research lab — shipped Sakana Fugu and Sakana Fugu Ultra. Both are models. Neither is a model in the sense you mean when you say "model." They are multi-agent orchestration systems delivered as a single API. You send a chat completion request. Fugu decides internally whether to answer directly, dispatch to a pool of expert agents, or recursively call itself.

On the benchmarks Sakana published today, Fugu Ultra beats the individual frontier models it orchestrates on 6 of 11 benchmarks, ties 1, and loses 4. The orchestrator is smarter than its workers. That result is the entire story.

The Architecture — Why This Isn't "Just a Router"

The first response from AI Twitter was "isn't this just a router?"

No. A router picks a model and forwards the call. Fugu is a trained coordinator that learns how to assign Thinker, Worker, and Verifier roles across a pool of frontier models, run multiple turns of agent collaboration, and synthesize the outputs into one answer. It can recursively call itself. It writes the prompts the agents use to talk to each other — those prompts are learned, not hand-written.

It is built on two ICLR 2026 papers from Sakana: TRINITY, an evolved coordinator that assigns roles across turns and adapts team composition per task; and The Conductor, a reinforcement-learned coordinator that discovers natural-language coordination strategies. Technical report on arXiv (2606.21228).

For builders: you call one endpoint and get multi-agent performance without writing a multi-agent framework. The API is OpenAI-compatible. Change the base URL and the model name to fugu-ultra-20260615 or fugu, and your prompts route through Sakana's coordinator.

The Numbers

Sakana published a full benchmark suite against the three frontier models in its pool. Fugu Ultra leads on most rows:

Bold = top score. Fugu Ultra wins 6, ties 1, loses 4 — the right shape for an orchestrator. It wins by routing work to whichever worker is best at that exact thing and synthesizing across them. The places it loses are the places where one worker has a domain edge: long-context retrieval is GPT-5.5's home turf, scientific code is Gemini's, and Fugu's recursion overhead does not pay off there. Regular Fugu (lower latency, opt-out available) also leads on SciCode (60.1), τ³ Banking (21.7), and Long Context (74.7) — lighter coordination wins on those.

Why This Is The Big Story

1. The "Use The Best Model" Era Is Over

For three years the dominant strategic question for builders has been which model should I default to? GPT-5.5 or Opus 4.8 or Gemini 3.1 Pro?

Fugu's answer is none of the above. You call Fugu. Fugu decides. You do not pick a model. You pick a policy — opt out the agents that violate your data residency rules, keep the ones that match your cost and quality bar — and let the orchestrator handle the rest.

That is a product surface change, not a benchmark change. The model layer is becoming commodity. The orchestration layer is becoming the product.

2. The Export Control Arbitrage Is Real

Anthropic's Fable 5 and Mythos 5 have been offline since June 12 under a US executive order. Ten days later Sakana — a Japanese lab — ships a model whose marketing line is "delivering frontier capability without the risk of export controls," claiming performance "shoulder-to-shoulder with Fable 5 and Mythos Preview." Neither banned model is in Fugu's pool because neither is publicly accessible. Fugu is matching them with what is publicly accessible.

The strategic question for any enterprise right now: do I want my AI capability subject to a US regulatory action I cannot predict? Fugu gives you a hedge today. Whether the hedge stays open is a policy question — but hedges are valuable precisely when you do not yet know whether you need them.

3. The Orchestrator Beats Its Workers

This is the research result that will be cited in papers for the next two years. The orchestrator outperforms the individual frontier models it orchestrates. Not by calling them in sequence and picking the best answer — by learning, through training, how to coordinate them.

The demos are the proof:

• AutoResearch. Fugu Ultra autonomously improved a small GPT's training recipe. 123 experiments over 14 hours on one H100. Best mean validation bits-per-byte: 0.9774. Frontier baselines: 0.9781, 0.9793, 0.9822. The orchestrator found a better recipe than the models it called could find on their own.
• Rubik's Cube solver. Pure Python, no libraries. Fugu Ultra solved 300/300 held-out cubes at 19.72 moves. Two of three frontier baselines crashed and solved 0/300.
• Classical Japanese kana reading order. On a 1610 letter by Hōshun'in, Fugu Ultra scored NED 0.80. The nearest frontier baseline scored 0.24.
• Blindfold chess. Four games from memory, no board shown. Fugu beat three frontier models and a 2100-Elo Stockfish. Won all four.

The orchestrator is doing real cognitive work, not just brokering API calls.

The Skeptic's Case

Three legitimate concerns. One: self-reported benchmarks. Sakana ran the tests, the model pool is undisclosed. Treat the numbers as plausible until an independent third party reproduces them. Two: routing opacity. You do not know which models handled which part of your prompt, and Fugu Ultra's pool is fixed with no opt-out — if you have compliance constraints, you get the regular tier and lose the headline performance. Three: the "wrapper" critique. Half the public reaction (6 of 12 posts Sakana reviewed) is skeptical. Demos argue for real orchestration intelligence. Closed pool argues for caution. Run your own evals.

What Builders Should Do This Week

1. Get a Sakana console account and call Fugu Ultra on your five hardest internal tasks. The OpenAI-compatible API means you swap it in for GPT-5.5 or Opus 4.8 in ten minutes. You will know within an afternoon whether the headline numbers translate to your workload.

2. Stop treating "model" as the strategic decision. Treat "model policy" as the strategic decision. Which models are in your pool? Which are opted out for compliance? What is your default for chat, coding, long-context retrieval? Fugu is the first product that makes model policy a first-class surface.

3. Audit your Fable 5 dependency. If you have built on Fable 5 or Mythos 5 and they are still offline on July 8, Fugu Ultra is the most credible drop-in. Evaluate it now, not under production pressure.

The Take

The frontier model race has been a race to build the biggest model. Sakana just won a different race by declining to enter that one. They built the best user of models. And the best user of models beats the best models.

That is not a benchmark win. That is a thesis. The model is the raw material, not the product. The product is the orchestration layer above it. The model-as-product era is closing. The orchestration-as-product era opened this week. The scoreboards have already changed.

— Mr. Technology