Sakana's new model beats the frontier by not using it

The posts

On September 11, @SakanaAILabs announced two models, Fugu Max and Fugu Ultra v2, calling them "the next evolution of Sakana Fugu's multi-agent orchestration system." Twenty-six minutes later, co-founder David Ha posted the reasoning behind the release from @hardmaru, and the reasoning is the more interesting half.

Fugu is not a model the way Claude or GPT is a model. It is a system that takes your request, hands pieces of it to a pool of other models, then coordinates and assembles the result. You call one endpoint. Several models do the work behind it.

Sakana is leading with the scores. It says Fugu Ultra v2 gets 48.3 on Chartography, a chart-reading and visual reasoning test, against 27.3 for Claude Opus 5 and 29.5 for Claude Fable 5. On DeepSWE, a software engineering benchmark, it reports 74.3.

The stranger detail is what is missing. Sakana says Fable 5, Fable 5.1 and GPT-6 Astra are not in Ultra v2's pool at all. The system reportedly outscoring those models does not have access to them. The pool is open weights and specialized models, including Nvidia's Nemotron family.

The pitch is about dependency, not scores

@hardmaru's post spends more words on risk than on benchmarks. "Relying on a single company's model for critical infrastructure is a massive risk," he wrote, citing export controls as evidence that access "can disappear overnight." His argument is that an orchestration layer routes around vendor restrictions because the pool underneath is swappable.

That argument lands harder this week than it would have in June. OpenAI's new Agents API arrived with US-only data residency written into its own documentation. Anthropic published a threat report accusing two Chinese AI apps of relaying live customer prompts to Claude. Terms of service are moving faster than most companies' integrations.

My read

The obvious objection to Fugu is that an orchestrator's score belongs to the system, not to anything inside it. Sakana did not build a model that outthinks Opus 5. It built a dispatcher that assembled a better answer out of cheaper parts. That is a real product and a genuinely good result, but it is a different claim than a leaderboard row suggests.

A few things temper the announcement:

• The benchmark numbers are vendor-reported. No independent evaluation has been published.
• Routing is undisclosed. You cannot see which model handled your request, which matters if any of your data is sensitive.
• You cannot self-host it. Access runs through Sakana's API only, despite the open-weights pool.
• Service is unavailable in the EU and EEA, which is an odd footnote under a sovereignty pitch.
• Ultra v2 costs $5 per million input tokens and $30 per million output, and both roughly double above 272K of context. Fugu Max is the cheap tier at $2 and $6.

So the hedge against vendor concentration is itself a single vendor, with less visibility into what happens to your prompt than you get from Anthropic or OpenAI directly. Worth looking at, worth reading the terms first.

What this means if your business runs on AI

Most small companies I work with in the Fox Valley did not choose an AI vendor. They chose a tool, and the tool chose the vendor. The quoting assistant and the thing that drafts follow-up emails are each wired to a specific API under a contract nobody on staff has read.

That was fine when the only question was quality. It is less fine now that residency rules and regional availability change every few weeks. Forget which model is smartest for a minute and ask what happens to your Tuesday if the model you depend on gets restricted or pulled out of your region.

The practical version of Sakana's argument does not require buying Fugu. Know where your automations actually call out, and build them so swapping the model underneath is a config change rather than a rebuild. It is boring architecture work, and it decides whether a vendor announcement is news or a fire drill.

If you want a clear picture of where your own processes depend on a single outside service, New Face Design does a free process audit. We map what you run, where it calls out, and what would break if one of those calls stopped answering. You keep the map whether or not you hire us.