Sakana launches Fugu Max to route requests across specialist models
Sakana says Fugu Max dynamically routes requests across open-weight and specialist models at two to six times lower cost than elite models. In the same release, the company says Fugu Ultra v2 led five of eight hard evaluation suites.
TL;DR
- Fugu Max and Fugu Ultra v2 divide Sakana Fugu between cost and peak performance, with SakanaAILabs' announcement describing Max as the outward move on the cost-capability frontier and Ultra as the upward one.
- Fugu Max is priced at $2 per million input tokens and $6 per million output tokens, according to SakanaAILabs' pricing post, which also claims performance within striking distance of elite models at two to six times lower cost.
- Fugu Ultra v2 reported a 74.3 DeepSWE score and 48.3 on Chartography, versus 27.3 for Opus 5, in SakanaAILabs' benchmark post.
- Sakana frames the swappable agent pool as protection against vendor lock-in and service cutoffs in hardmaru's launch post.
Sakana's release post says users already on Fugu can select Max or Ultra v2 with a one-line parameter change. The official Max benchmark chart places its claimed wins alongside single models in a similar price range.
One architecture, two operating points
Sakana says Max and Ultra v2 use the same orchestration architecture, optimized for different objectives. Max routes toward the least expensive capable option; Ultra v2 is tuned for complex, multi-step work where raw output quality takes priority.
Fugu Max routing and price
Fugu Max expands the model pool with open-weight and specialist models, including NVIDIA's Nemotron family. Sakana says the router selects the leanest model capable of a task rather than sending every request to one fixed model.
- Price: $2 per million input tokens and $6 per million output tokens.
- Relative pricing claim: 40% to 60% lower output pricing than Sonnet 5, GPT 5.6 Terra, and Kimi K3, according to Sakana's release post.
- Reported wins: best overall score on Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. Sakana identifies SWEFish as its own coding-challenge evaluation.
- Frontier claim: a cost-performance efficient position on seven of ten benchmarks.
Fugu Ultra v2 benchmark claims
Ultra v2 is Sakana's higher-capability setting. The company reports best or joint-best results on five of eight suites, GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon, and a top-two placement on seven.
- Chartography: 48.3, compared with Opus 5 at 27.3 and Fable 5 at 29.5 on Sakana's published Ultra v2 chart.
- DeepSWE: 74.3. Sakana says it exceeds models costing three to five times more per token.
Training cutoff and pool exclusions
Sakana lists an August 28, 2026 training cutoff for Ultra v2 and says Fable 5, Fable 5.1, and GPT-6-Astra are absent from its agent pool. The release post names those exclusions but does not publish a complete model-and-version inventory for the remaining pool.