Sakana Fugu Max routes to the leanest model — 40-60% under frontier
A learned router behind one OpenAI-compatible API: Max undercuts frontier pricing on six benchmarks; Ultra v2 beats Opus 5 on charts. Plus an open Suno rival.
Copy markdown
One API, a learned router underneath
Fugu Max and Fugu Ultra v2 aren't single models — they're learned orchestrators behind one OpenAI-compatible API that route each task to the leanest model that can solve it, now across a pool of open-weight and specialized models (Sakana built it out with NVIDIA's Nemotron family). Both are live today, and existing Fugu users switch with a one-line parameter change.
$2/$6, and it tops six benchmarks
Fugu Max is $2 per million input tokens and $6 per million output — output pricing 40-60% below Sonnet 5, GPT-5.6 Terra, and Kimi K3. It posts the best overall score on six benchmarks (Terminal Bench 2.1, GPQAD, AA-LCR, SWEFish and more) and extends the cost-performance frontier on 7 of 10.
Fugu Ultra v2 for the work Max can't touch
The capability tier scores 48.3 on Chartography against Opus 5's 27.3 and Fable 5's 29.5, plus 74.3 on DeepSWE, taking best or joint-best on five of eight benchmarks. Sakana aims it at autonomous research and full-stack builds you'd otherwise hand to a frontier model.
The real win: you stop betting on one lab
Because neither tier leans on a single proprietary model, when GPT-6 Astra or Fable 5 moves on price or availability it's Sakana's routing problem, not your rewrite. Build this weekend: point your existing OpenAI SDK at Fugu's base URL and A/B your current model against Max on live traffic — it's a one-line change, so the cost delta shows up within a day.
Elsewhere: an open music model that beats Suno
M-A-P released YuE2-3B, a 7.3 GB open-weights model that generates full songs — lyrics, vocals and backing — and reportedly outscores Suno v5 and v6 on its makers' own WildSongBench. Weights and a GGUF build are already on Hugging Face, giving you a local, license-clean alternative to run yourself.