Curated Pool

Fugu Ultra v2.0 is not a model. It's a claim laundered through a benchmark. Sakana AI shipped it September 11, alongside Fugu Max, and the number everyone will repeat by end of week is 48.3 on Chartography, nearly double Opus 5's 27.3. That is a genuinely large gap. It is also a gap measured against a pool that, per the same reporting, excludes Fable 5 and GPT-6-Astra. A second source fills in one of those blanks and finds Fable 5 at 29.5 — still nowhere near 48.3, true, but that's not the point. The point is the victory lap was written before the competitors were in the room.

Eleven models in eleven days from seven providers is not an industry, it's a scroll feed, and Fugu Ultra v2 is engineered to win the scroll. It's not a base model at all — Sakana calls it an "orchestration" layer, and Fugu Max's entire pitch is that it now conducts more of the field, open-weights and specialized checkpoints alike, including NVIDIA's Nemotron family via a new tie-up. That's a legitimate architectural bet: don't train the smartest model, train the smartest referee. But a referee's scorecard is only as honest as the players it invited, and right now the invite list is curated.

The pricing is the tell that this is a business decision dressed as a research one. Fugu Ultra v2 runs $5 in, $30 out per million tokens on a million-token window, per OpenRouter's listing, while Fugu Max undercuts Sonnet 5, GPT 5.6 Terra, and Kimi K3 on output cost by 40 to 60 percent, by Sakana's own account. You don't cut prices that hard on a model you're confident wins on merit alone; you cut prices to buy adoption while the benchmark story is still fresh and uncontested.

Second wind, dropped — that's the whole genre this release belongs to. Sakana isn't debuting a new capability, it's re-launching an old thesis (models that orchestrate other models) with a bigger number and a smaller comparison set, timed to land before Fable 5 and GPT-6-Astra get their own turn at the same test. The 74.3 DeepSWE score is real and worth noting on its own terms. The Chartography number is a magic trick performed with half the deck missing. Call it what it is when the full pool shows up.

Nine days into September, the openrouter listing for Sakana AI's newest release reads like a chart drop sheet: two models, same day, same lineage. Fugu Ultra v2.0 and Fugu Max land together on September 11, 2026, the tenth and eleventh entries in a month that has already seen output from seven different labs — a release cadence closer to a mixtape season than a product roadmap. Fugu Ultra v2 isn't a lone drop; it's a week's worth of shelf space claimed at once, Max positioned as the accessible pressing at $2 per million input tokens and $6 per million output, Ultra v2 held back as the flagship cut.

What's worth sitting with isn't the release itself but the way the numbers were staged. Pondero's initial write-up clocked Fugu Ultra v2 at 48.3 on Chartography, a visual-reasoning benchmark, against Opus 5's 27.3 — a gap wide enough to make headlines — but the pool it was measured against was missing two names, Fable 5 and GPT-6-Astra, both established enough that their absence reads less like oversight than staging. Three days later, ai-tldr's rerun folds Fable 5 back in at 29.5, still well behind Ultra v2, and adds a second axis entirely: 74.3 on DeepSWE, the software-engineering benchmark, a number nobody was citing on release day. The story shifts from "beat Opus 5" to "beat almost everyone, on almost everything" — top two on seven of eight benchmarks, including GDP.pdf, Chartography, DeepSWE, Toolathon, and Sakana's own SWEFish.

That drift — from a curated first cut to a fuller accounting three days out — is the actual texture of this season's model culture. Benchmarks are being released serially now, the way a scene rolls out a single, then a video, then a remix with a feature nobody expected. The comparison set isn't neutral data; it's a track listing, assembled and then revised once the room notices who's missing. Eleven models from seven providers in eleven days means nobody's citation is stable for more than seventy-two hours — which is less a benchmarking problem than a distribution problem, the same one music scenes hit when everyone drops on the same Friday and the algorithm decides whose numbers get amplified before the corrections arrive.