Sakana Fugu and the Routing Stack of 2026

!Three routing layers — orchestration model, API gateway, coding agent

On June 22, 2026, the Tokyo lab Sakana AI shipped a product that is easy to describe and hard to place. It is called Fugu, after the pufferfish, and it is a language model whose job is to call other language models. You send one request to one endpoint. Behind the endpoint, Fugu decides whether to answer the question itself or assemble a team of frontier models — GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro, open-weight models, and, recursively, instances of itself — and then synthesizes their work into a single reply.

Within a day, the comparisons arrived: Is this just OpenRouter? Is this Kilo Code with a different label? The questions are reasonable and the answer is no, but the reason it is no turns out to be the most useful thing to understand about the current moment in applied AI. Fugu, OpenRouter, and Kilo Code all sit between you and the models. They all route. But they route at three different altitudes, and conflating them hides the actual decision a buyer has to make.

This piece does three things. First, it separates the three layers and shows what each one genuinely decides on your behalf. Second, it lays out the verified Fugu evidence — the benchmarks Sakana published, the prices, and the early independent testing, which tells a noticeably cooler story than the launch post. Third, it offers a decision framework, because the right answer depends almost entirely on which problem you actually have.

Three Things That All Call Themselves Routers

The word "routing" has quietly come to mean three unrelated activities. Pulling them apart is the whole game.

The distinction that matters: who holds the intelligence. OpenRouter is plumbing — it pools 400+ models across 70+ providers behind one OpenAI-compatible endpoint, passes through provider pricing, and keeps your app from returning a 500 when Anthropic rate-limits you mid-traffic. It does not try to be smart about your prompt unless you opt into its Auto Router, and even then the routing is a classifier choosing one model, not a coordinator running several. Kilo Code is a workspace — an open-source agent (Apache 2.0 extension, 420K+ installs) that brings five specialist modes into VS Code and JetBrains, lets you pick a model per mode from 500+ options, and in Orchestrator mode breaks a task into subtasks routed to the right specialist. The intelligence about how to build software lives in Kilo's scaffolding and in you.

Fugu moves that intelligence into a model. It is not a config file and not an editor. It is an LLM that learned — through large-scale fine-tuning, evolutionary search, and reinforcement learning, building on Sakana's two ICLR 2026 papers, Trinity and Conductor — to construct an "agentic scaffold" on the fly: pick the workers, assign the roles, run the verification passes, merge the answers. The scaffold is generated per query rather than prescribed by a human. That is the genuinely new idea, and it is also the source of every problem the early testers ran into.

What Actually Launched

There is a notable absence in the pool. Anthropic's Fable 5 and Mythos are not among the models Fugu can call, because they are no longer publicly accessible — the very restriction that motivated the product. So Sakana's claim that Fugu "stands shoulder-to-shoulder with Fable and Mythos" is a claim about matching models it cannot use, which is either the most impressive part of the story or a rhetorical sleight, depending on how the benchmarks hold up.

The Benchmarks Sakana Published

These are vendor-reported numbers. No independent replication exists two days after launch, and the comparison columns notably exclude the restricted Anthropic models Fugu claims to rival.

Read carefully, two patterns emerge. The coding leads are large and would matter if they survive contact with an independent harness: Fugu Ultra's 73.7 on SWE-Bench Pro is 4.5 points above Opus 4.8 and roughly 15 above GPT-5.5. The reasoning and long-context leads are thin — a point here, a few points there — and on MRCRv2, the long-context retrieval test, GPT-5.5 still wins outright. An orchestrator that runs several strong models and keeps the best answer should edge out any single member of its pool. The interesting question is not whether it wins on paper but what that win costs in time and money, which the benchmark table is silent about.

The Cost, Including the Part Nobody Advertises

Sakana sells Fugu two ways.

Subscriptions (both models available in every tier; the difference is usage volume):

Pay-as-you-go (Fugu Ultra, per 1M tokens):

For base Fugu, the billing has a genuinely clever wrinkle: when Fugu fires several agents on one request, you are not charged a stacked fee per agent. You pay a single rate based on the top-tier model involved. That is the headline mechanic, and it is real.

The wrinkle that is not in the headline lives in the usage object. Fugu Ultra returns orchestration_input_tokens and orchestration_output_tokens separately from the visible input and output. Unlike the way OpenAI accounts for reasoning tokens, these orchestration tokens are real charges — billed at the same rate as standard tokens, on top of the work you can see. The coordinator's internal deliberation, the worker hand-offs, the verification passes: all of it is metered. This is the line item that turns a clean per-token price into something you cannot estimate from the prompt alone, and it is the mechanism behind the most alarming early report — a tester who burned his entire five-hour quota on a single prompt.

The Hands-On Reality Check

This is the part of the story the launch post does not tell, and it is the part that should govern any near-term decision. All of the following are named testers, June 22–23.

The pattern is consistent enough to be useful. Fugu is strong on long, messy, multi-step analytical work — code review, patent-landscape mapping, literature synthesis — where running several models and reconciling them genuinely beats one model thinking alone. It is weak and slow on interactive coding, especially frontend and graphics, where latency and fast iteration matter more than the marginal quality of any single pass. The Santos result is the cleanest illustration of the trade the whole product makes: meaningfully faster and cheaper than a premium single model on a bounded task, with output a discerning user liked slightly less.

The Critique That Survives Scrutiny

Sakana's framing leans hard on sovereignty and resilience — the idea that an orchestrator over a swappable pool protects you from any one provider disappearing. The logic has a hole that critics found within hours. Fugu's capability is entirely derived from the models in its pool. If the frontier providers are restricted simultaneously — the scenario export controls actually produce — Fugu's options shrink in lockstep. Diversification across vendors is real protection against one vendor having an outage or a price hike. It is not sovereignty, because Sakana controls none of the intelligence it resells. The company benchmarking against models it cannot legally include in its own pool is the same tension viewed from the other side.

This is where the layer distinction pays off. OpenRouter makes the identical resilience promise and is honest about its altitude: it is failover plumbing, it never claims the routing is the product, and you keep provider-level pricing with no markup beyond a platform fee on paid tiers. Kilo Code makes a different honesty available — it is open source, BYOK, zero-markup, and it shows you exactly which model ran which subtask. Fugu's opacity is the cost of its ambition: it will not tell you which model answered, which is fine until you need to debug a regression or reproduce a result.

A Decision Framework

The three tools are not substitutes. Map the problem to the layer.

The honest summary: if you want plumbing, use OpenRouter; if you want a workspace, use Kilo Code; Fugu is competing with neither — it is competing with the idea that you should pick a single frontier model at all, and it wins that argument only on a specific class of slow, deep, analytical tasks.

What Happens Next

Three predictions, weighted by the evidence.

The binding constraint on Fugu is latency and quota, not quality. Thirty-minute runs and single-prompt quota exhaustion are adoption-killers for interactive work, and pricing opacity around orchestration tokens is the second-loudest complaint. Expect Sakana to ship a faster routing path and clearer cost reporting within a quarter, and EU availability once the GDPR work lands. These are fixable and the company knows exactly which numbers are hurting it.

The structural bet is more interesting than the product. Ha is arguing that orchestration models compound automatically — Fugu improves whenever any model in its pool improves, without Sakana training anything new. That is a real position in a field exhausted by the cost of training-scale arms races, and if routing quality keeps climbing it is a defensible one. It also reframes OpenRouter and Kilo Code not as competitors but as the layers a mature stack will run underneath an orchestration model: the gateway handles provider failover, the agent handles the editor surface, and a learned coordinator sits on top deciding what to call.

The risk is the same lock-in Fugu claims to cure. The entire model depends on frontier vendors permitting their APIs to be used as silent sub-agents inside a competitor's product. If OpenAI, Google, or Anthropic tighten terms against being resold as anonymous workers — and an orchestrator that hides which model it used gives them reason to — the pool model faces a commercial squeeze that no amount of clever routing can engineer around. The bet that you can build a durable business on top of models you do not control is exactly the bet Fugu asks its customers to stop making.

For now, the verdict is narrower than the launch suggested and more interesting than the skeptics allowed. Fugu is not a Claude or GPT replacement, and it is not an OpenRouter or Kilo Code replacement. It is the first serious attempt to make "which model should answer this" a learned decision rather than a configured one — and the early evidence says that decision is worth paying for on deep analytical work, and not yet worth its latency anywhere speed matters.

---

Sources

- sakana.ai/fugu-release — primary launch announcement - console.sakana.ai — pricing and model pages (primary) - arXiv:2606.21228 — Sakana Fugu Technical Report - openrouter.ai/blog/insights/model-routing — how OpenRouter routing works - openrouter.ai/pricing — OpenRouter pricing tiers - kilo.ai/articles/coding-agents-for-vscode — Kilo Code capabilities and modes - blog.kilo.ai — Kilo Code VS Code rebuild notes - VentureBeat (Jun 22, 2026) - Nikkei Asia (Jun 23, 2026) - The Verge (Jun 23, 2026) - the-decoder (Jun 23, 2026) - Business Standard (Jun 23, 2026) - MarkTechPost (Jun 22, 2026) - DevelopersIO / Classmethod hands-on (Jun 22, 2026) - Named tester posts from Ethan Mollick, Peter Steinberger, Hamel Husain, @LLMJunky, and Mark Santos

All Fugu benchmark figures are vendor-reported. No independent replication existed at the time of writing.

Last updated: June 24, 2026